portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-06-17 15:27:06 +01:00
Code Releases Activity

Merge branch 'develop' into feature/gpt

Browse Source
This commit is contained in:
Christoph Lehner
2020-05-06 15:03:35 +02:00
committed by GitHub
parent 87984ece7d 224cbf0453
commit 3c6ffcb48c
192 changed files with 3920 additions and 6302 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
Grid/algorithms/Algorithms.h
View File

@ -39,14 +39,18 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
#include <Grid/algorithms/approx/Remez.h>
#include <Grid/algorithms/approx/MultiShiftFunction.h>
#include <Grid/algorithms/approx/Forecast.h>
#include <Grid/algorithms/approx/RemezGeneral.h>
#include <Grid/algorithms/approx/ZMobius.h>
#include <Grid/algorithms/iterative/Deflation.h>
#include <Grid/algorithms/iterative/ConjugateGradient.h>
#include <Grid/algorithms/iterative/BiCGSTAB.h>
#include <Grid/algorithms/iterative/ConjugateResidual.h>
#include <Grid/algorithms/iterative/NormalEquations.h>
#include <Grid/algorithms/iterative/SchurRedBlack.h>
#include <Grid/algorithms/iterative/ConjugateGradientMultiShift.h>
#include <Grid/algorithms/iterative/ConjugateGradientMixedPrec.h>
#include <Grid/algorithms/iterative/BiCGSTABMixedPrec.h>
#include <Grid/algorithms/iterative/BlockConjugateGradient.h>
#include <Grid/algorithms/iterative/ConjugateGradientReliableUpdate.h>
#include <Grid/algorithms/iterative/MinimalResidual.h>

126
Grid/algorithms/LinearOperator.h
View File

@ -334,6 +334,132 @@ public:
return axpy_norm(out,-1.0,tmp,in);
}
};
template<class Field>
class NonHermitianSchurOperatorBase : public LinearOperatorBase<Field>
{
public:
virtual RealD Mpc (const Field& in, Field& out) = 0;
virtual RealD MpcDag (const Field& in, Field& out) = 0;
virtual void MpcDagMpc(const Field& in, Field& out, RealD& ni, RealD& no) {
Field tmp(in.Grid());
tmp.Checkerboard() = in.Checkerboard();
ni = Mpc(in,tmp);
no = MpcDag(tmp,out);
}
virtual void HermOpAndNorm(const Field& in, Field& out, RealD& n1, RealD& n2) {
assert(0);
}
virtual void HermOp(const Field& in, Field& out) {
assert(0);
}
void Op(const Field& in, Field& out) {
Mpc(in, out);
}
void AdjOp(const Field& in, Field& out) {
MpcDag(in, out);
}
// Support for coarsening to a multigrid
void OpDiag(const Field& in, Field& out) {
assert(0); // must coarsen the unpreconditioned system
}
void OpDir(const Field& in, Field& out, int dir, int disp) {
assert(0);
}
};
template<class Matrix, class Field>
class NonHermitianSchurDiagMooeeOperator : public NonHermitianSchurOperatorBase<Field>
{
public:
Matrix& _Mat;
NonHermitianSchurDiagMooeeOperator(Matrix& Mat): _Mat(Mat){};
virtual RealD Mpc(const Field& in, Field& out) {
Field tmp(in.Grid());
tmp.Checkerboard() = !in.Checkerboard();
_Mat.Meooe(in, tmp);
_Mat.MooeeInv(tmp, out);
_Mat.Meooe(out, tmp);
_Mat.Mooee(in, out);
return axpy_norm(out, -1.0, tmp, out);
}
virtual RealD MpcDag(const Field& in, Field& out) {
Field tmp(in.Grid());
_Mat.MeooeDag(in, tmp);
_Mat.MooeeInvDag(tmp, out);
_Mat.MeooeDag(out, tmp);
_Mat.MooeeDag(in, out);
return axpy_norm(out, -1.0, tmp, out);
}
};
template<class Matrix,class Field>
class NonHermitianSchurDiagOneOperator : public NonHermitianSchurOperatorBase<Field>
{
protected:
Matrix &_Mat;
public:
NonHermitianSchurDiagOneOperator (Matrix& Mat): _Mat(Mat){};
virtual RealD Mpc(const Field& in, Field& out) {
Field tmp(in.Grid());
_Mat.Meooe(in, out);
_Mat.MooeeInv(out, tmp);
_Mat.Meooe(tmp, out);
_Mat.MooeeInv(out, tmp);
return axpy_norm(out, -1.0, tmp, in);
}
virtual RealD MpcDag(const Field& in, Field& out) {
Field tmp(in.Grid());
_Mat.MooeeInvDag(in, out);
_Mat.MeooeDag(out, tmp);
_Mat.MooeeInvDag(tmp, out);
_Mat.MeooeDag(out, tmp);
return axpy_norm(out, -1.0, tmp, in);
}
};
template<class Matrix, class Field>
class NonHermitianSchurDiagTwoOperator : public NonHermitianSchurOperatorBase<Field>
{
protected:
Matrix& _Mat;
public:
NonHermitianSchurDiagTwoOperator(Matrix& Mat): _Mat(Mat){};
virtual RealD Mpc(const Field& in, Field& out) {
Field tmp(in.Grid());
_Mat.MooeeInv(in, out);
_Mat.Meooe(out, tmp);
_Mat.MooeeInv(tmp, out);
_Mat.Meooe(out, tmp);
return axpy_norm(out, -1.0, tmp, in);
}
virtual RealD MpcDag(const Field& in, Field& out) {
Field tmp(in.Grid());
_Mat.MeooeDag(in, out);
_Mat.MooeeInvDag(out, tmp);
_Mat.MeooeDag(tmp, out);
_Mat.MooeeInvDag(out, tmp);
return axpy_norm(out, -1.0, tmp, in);
}
};
///////////////////////////////////////////////////////////////////////////////////////////////////
// Left handed Moo^-1 ; (Moo - Moe Mee^-1 Meo) psi = eta --> ( 1 - Moo^-1 Moe Mee^-1 Meo ) psi = Moo^-1 eta
// Right handed Moo^-1 ; (Moo - Moe Mee^-1 Meo) Moo^-1 Moo psi = eta --> ( 1 - Moe Mee^-1 Meo Moo^-1) phi=eta ; psi = Moo^-1 phi

473
Grid/algorithms/approx/RemezGeneral.cc Normal file
View File

@ -0,0 +1,473 @@
#include<math.h>
#include<stdio.h>
#include<stdlib.h>
#include<string>
#include<iostream>
#include<iomanip>
#include<cassert>
#include<Grid/algorithms/approx/RemezGeneral.h>
// Constructor
AlgRemezGeneral::AlgRemezGeneral(double lower, double upper, long precision,
bigfloat (*f)(bigfloat x, void *data), void *data): f(f),
data(data),
prec(precision),
apstrt(lower), apend(upper), apwidt(upper - lower),
n(0), d(0), pow_n(0), pow_d(0)
{
bigfloat::setDefaultPrecision(prec);
std::cout<<"Approximation bounds are ["<<apstrt<<","<<apend<<"]\n";
std::cout<<"Precision of arithmetic is "<<precision<<std::endl;
}
//Determine the properties of the numerator and denominator polynomials
void AlgRemezGeneral::setupPolyProperties(int num_degree, int den_degree, PolyType num_type_in, PolyType den_type_in){
pow_n = num_degree;
pow_d = den_degree;
if(pow_n % 2 == 0 && num_type_in == PolyType::Odd) assert(0);
if(pow_n % 2 == 1 && num_type_in == PolyType::Even) assert(0);
if(pow_d % 2 == 0 && den_type_in == PolyType::Odd) assert(0);
if(pow_d % 2 == 1 && den_type_in == PolyType::Even) assert(0);
num_type = num_type_in;
den_type = den_type_in;
num_pows.resize(pow_n+1);
den_pows.resize(pow_d+1);
int n_in = 0;
bool odd = num_type == PolyType::Full || num_type == PolyType::Odd;
bool even = num_type == PolyType::Full || num_type == PolyType::Even;
for(int i=0;i<=pow_n;i++){
num_pows[i] = -1;
if(i % 2 == 0 && even) num_pows[i] = n_in++;
if(i % 2 == 1 && odd) num_pows[i] = n_in++;
}
std::cout << n_in << " terms in numerator" << std::endl;
--n_in; //power is 1 less than the number of terms, eg pow=1 a x^1 + b x^0
int d_in = 0;
odd = den_type == PolyType::Full || den_type == PolyType::Odd;
even = den_type == PolyType::Full || den_type == PolyType::Even;
for(int i=0;i<=pow_d;i++){
den_pows[i] = -1;
if(i % 2 == 0 && even) den_pows[i] = d_in++;
if(i % 2 == 1 && odd) den_pows[i] = d_in++;
}
std::cout << d_in << " terms in denominator" << std::endl;
--d_in;
n = n_in;
d = d_in;
}
//Setup algorithm
void AlgRemezGeneral::reinitializeAlgorithm(){
spread = 1.0e37;
iter = 0;
neq = n + d + 1; //not +2 because highest-power term in denominator is fixed to 1
param.resize(neq);
yy.resize(neq+1);
//Initialize linear equation temporaries
A.resize(neq*neq);
B.resize(neq);
IPS.resize(neq);
//Initialize maximum and minimum errors
xx.resize(neq+2);
mm.resize(neq+1);
initialGuess();
//Initialize search steps
step.resize(neq+1);
stpini();
}
double AlgRemezGeneral::generateApprox(const int num_degree, const int den_degree,
const PolyType num_type_in, const PolyType den_type_in,
const double _tolerance, const int report_freq){
//Setup the properties of the polynomial
setupPolyProperties(num_degree, den_degree, num_type_in, den_type_in);
//Setup the algorithm
reinitializeAlgorithm();
bigfloat tolerance = _tolerance;
//Iterate until convergance
while (spread > tolerance) {
if (iter++ % report_freq==0)
std::cout<<"Iteration " <<iter-1<<" spread "<<(double)spread<<" delta "<<(double)delta << std::endl;
equations();
if (delta < tolerance) {
std::cout<<"Iteration " << iter-1 << " delta too small (" << delta << "<" << tolerance << "), try increasing precision\n";
assert(0);
};
assert( delta>= tolerance );
search();
}
int sign;
double error = (double)getErr(mm[0],&sign);
std::cout<<"Converged at "<<iter<<" iterations; error = "<<error<<std::endl;
// Return the maximum error in the approximation
return error;
}
// Initial values of maximal and minimal errors
void AlgRemezGeneral::initialGuess(){
// Supply initial guesses for solution points
long ncheb = neq; // Degree of Chebyshev error estimate
// Find ncheb+1 extrema of Chebyshev polynomial
bigfloat a = ncheb;
bigfloat r;
mm[0] = apstrt;
for (long i = 1; i < ncheb; i++) {
r = 0.5 * (1 - cos((M_PI * i)/(double) a));
//r *= sqrt_bf(r);
r = (exp((double)r)-1.0)/(exp(1.0)-1.0);
mm[i] = apstrt + r * apwidt;
}
mm[ncheb] = apend;
a = 2.0 * ncheb;
for (long i = 0; i <= ncheb; i++) {
r = 0.5 * (1 - cos(M_PI * (2*i+1)/(double) a));
//r *= sqrt_bf(r); // Squeeze to low end of interval
r = (exp((double)r)-1.0)/(exp(1.0)-1.0);
xx[i] = apstrt + r * apwidt;
}
}
// Initialise step sizes
void AlgRemezGeneral::stpini(){
xx[neq+1] = apend;
delta = 0.25;
step[0] = xx[0] - apstrt;
for (int i = 1; i < neq; i++) step[i] = xx[i] - xx[i-1];
step[neq] = step[neq-1];
}
// Search for error maxima and minima
void AlgRemezGeneral::search(){
bigfloat a, q, xm, ym, xn, yn, xx1;
int emsign, ensign, steps;
int meq = neq + 1;
bigfloat eclose = 1.0e30;
bigfloat farther = 0l;
bigfloat xx0 = apstrt;
for (int i = 0; i < meq; i++) {
steps = 0;
xx1 = xx[i]; // Next zero
if (i == meq-1) xx1 = apend;
xm = mm[i];
ym = getErr(xm,&emsign);
q = step[i];
xn = xm + q;
if (xn < xx0 || xn >= xx1) { // Cannot skip over adjacent boundaries
q = -q;
xn = xm;
yn = ym;
ensign = emsign;
} else {
yn = getErr(xn,&ensign);
if (yn < ym) {
q = -q;
xn = xm;
yn = ym;
ensign = emsign;
}
}
while(yn >= ym) { // March until error becomes smaller.
if (++steps > 10)
break;
ym = yn;
xm = xn;
emsign = ensign;
a = xm + q;
if (a == xm || a <= xx0 || a >= xx1)
break;// Must not skip over the zeros either side.
xn = a;
yn = getErr(xn,&ensign);
}
mm[i] = xm; // Position of maximum
yy[i] = ym; // Value of maximum
if (eclose > ym) eclose = ym;
if (farther < ym) farther = ym;
xx0 = xx1; // Walk to next zero.
} // end of search loop
q = (farther - eclose); // Decrease step size if error spread increased
if (eclose != 0.0) q /= eclose; // Relative error spread
if (q >= spread)
delta *= 0.5; // Spread is increasing; decrease step size
spread = q;
for (int i = 0; i < neq; i++) {
q = yy[i+1];
if (q != 0.0) q = yy[i] / q - (bigfloat)1l;
else q = 0.0625;
if (q > (bigfloat)0.25) q = 0.25;
q *= mm[i+1] - mm[i];
step[i] = q * delta;
}
step[neq] = step[neq-1];
for (int i = 0; i < neq; i++) { // Insert new locations for the zeros.
xm = xx[i] - step[i];
if (xm <= apstrt)
continue;
if (xm >= apend)
continue;
if (xm <= mm[i])
xm = (bigfloat)0.5 * (mm[i] + xx[i]);
if (xm >= mm[i+1])
xm = (bigfloat)0.5 * (mm[i+1] + xx[i]);
xx[i] = xm;
}
}
// Solve the equations
void AlgRemezGeneral::equations(){
bigfloat x, y, z;
bigfloat *aa;
for (int i = 0; i < neq; i++) { // set up the equations for solution by simq()
int ip = neq * i; // offset to 1st element of this row of matrix
x = xx[i]; // the guess for this row
y = func(x); // right-hand-side vector
z = (bigfloat)1l;
aa = A.data()+ip;
int t = 0;
for (int j = 0; j <= pow_n; j++) {
if(num_pows[j] != -1){ *aa++ = z; t++; }
z *= x;
}
assert(t == n+1);
z = (bigfloat)1l;
t = 0;
for (int j = 0; j < pow_d; j++) {
if(den_pows[j] != -1){ *aa++ = -y * z; t++; }
z *= x;
}
assert(t == d);
B[i] = y * z; // Right hand side vector
}
// Solve the simultaneous linear equations.
if (simq()){
std::cout<<"simq failed\n";
exit(0);
}
}
// Evaluate the rational form P(x)/Q(x) using coefficients
// from the solution vector param
bigfloat AlgRemezGeneral::approx(const bigfloat x) const{
// Work backwards toward the constant term.
int c = n;
bigfloat yn = param[c--]; // Highest order numerator coefficient
for (int i = pow_n-1; i >= 0; i--) yn = x * yn + (num_pows[i] != -1 ? param[c--] : bigfloat(0l));
c = n+d;
bigfloat yd = 1l; //Highest degree coefficient is 1.0
for (int i = pow_d-1; i >= 0; i--) yd = x * yd + (den_pows[i] != -1 ? param[c--] : bigfloat(0l));
return(yn/yd);
}
// Compute size and sign of the approximation error at x
bigfloat AlgRemezGeneral::getErr(bigfloat x, int *sign) const{
bigfloat f = func(x);
bigfloat e = approx(x) - f;
if (f != 0) e /= f;
if (e < (bigfloat)0.0) {
*sign = -1;
e = -e;
}
else *sign = 1;
return(e);
}
// Solve the system AX=B
int AlgRemezGeneral::simq(){
int ip, ipj, ipk, ipn;
int idxpiv;
int kp, kp1, kpk, kpn;
int nip, nkp;
bigfloat em, q, rownrm, big, size, pivot, sum;
bigfloat *aa;
bigfloat *X = param.data();
int n = neq;
int nm1 = n - 1;
// Initialize IPS and X
int ij = 0;
for (int i = 0; i < n; i++) {
IPS[i] = i;
rownrm = 0.0;
for(int j = 0; j < n; j++) {
q = abs_bf(A[ij]);
if(rownrm < q) rownrm = q;
++ij;
}
if (rownrm == (bigfloat)0l) {
std::cout<<"simq rownrm=0\n";
return(1);
}
X[i] = (bigfloat)1.0 / rownrm;
}
for (int k = 0; k < nm1; k++) {
big = 0.0;
idxpiv = 0;
for (int i = k; i < n; i++) {
ip = IPS[i];
ipk = n*ip + k;
size = abs_bf(A[ipk]) * X[ip];
if (size > big) {
big = size;
idxpiv = i;
}
}
if (big == (bigfloat)0l) {
std::cout<<"simq big=0\n";
return(2);
}
if (idxpiv != k) {
int j = IPS[k];
IPS[k] = IPS[idxpiv];
IPS[idxpiv] = j;
}
kp = IPS[k];
kpk = n*kp + k;
pivot = A[kpk];
kp1 = k+1;
for (int i = kp1; i < n; i++) {
ip = IPS[i];
ipk = n*ip + k;
em = -A[ipk] / pivot;
A[ipk] = -em;
nip = n*ip;
nkp = n*kp;
aa = A.data()+nkp+kp1;
for (int j = kp1; j < n; j++) {
ipj = nip + j;
A[ipj] = A[ipj] + em * *aa++;
}
}
}
kpn = n * IPS[n-1] + n - 1; // last element of IPS[n] th row
if (A[kpn] == (bigfloat)0l) {
std::cout<<"simq A[kpn]=0\n";
return(3);
}
ip = IPS[0];
X[0] = B[ip];
for (int i = 1; i < n; i++) {
ip = IPS[i];
ipj = n * ip;
sum = 0.0;
for (int j = 0; j < i; j++) {
sum += A[ipj] * X[j];
++ipj;
}
X[i] = B[ip] - sum;
}
ipn = n * IPS[n-1] + n - 1;
X[n-1] = X[n-1] / A[ipn];
for (int iback = 1; iback < n; iback++) {
//i goes (n-1),...,1
int i = nm1 - iback;
ip = IPS[i];
nip = n*ip;
sum = 0.0;
aa = A.data()+nip+i+1;
for (int j= i + 1; j < n; j++)
sum += *aa++ * X[j];
X[i] = (X[i] - sum) / A[nip+i];
}
return(0);
}
void AlgRemezGeneral::csv(std::ostream & os) const{
os << "Numerator" << std::endl;
for(int i=0;i<=pow_n;i++){
os << getCoeffNum(i) << "*x^" << i;
if(i!=pow_n) os << " + ";
}
os << std::endl;
os << "Denominator" << std::endl;
for(int i=0;i<=pow_d;i++){
os << getCoeffDen(i) << "*x^" << i;
if(i!=pow_d) os << " + ";
}
os << std::endl;
//For a true minimax solution the errors should all be equal and the signs should oscillate +-+-+- etc
int sign;
os << "Errors at maxima: coordinate, error, (sign)" << std::endl;
for(int i=0;i<neq+1;i++){
os << mm[i] << " " << getErr(mm[i],&sign) << " (" << sign << ")" << std::endl;
}
os << "Scan over range:" << std::endl;
int npt = 60;
bigfloat dlt = (apend - apstrt)/bigfloat(npt-1);
for (bigfloat x=apstrt; x<=apend; x = x + dlt) {
double f = evaluateFunc(x);
double r = evaluateApprox(x);
os<< x<<","<<r<<","<<f<<","<<r-f<<std::endl;
}
return;
}

170
Grid/algorithms/approx/RemezGeneral.h Normal file
View File

@ -0,0 +1,170 @@
/*
C.Kelly Jan 2020 based on implementation by M. Clark May 2005
AlgRemezGeneral is an implementation of the Remez algorithm for approximating an arbitrary function by a rational polynomial
It includes optional restriction to odd/even polynomials for the numerator and/or denominator
*/
#ifndef INCLUDED_ALG_REMEZ_GENERAL_H
#define INCLUDED_ALG_REMEZ_GENERAL_H
#include <stddef.h>
#include <Grid/GridStd.h>
#ifdef HAVE_LIBGMP
#include "bigfloat.h"
#else
#include "bigfloat_double.h"
#endif
class AlgRemezGeneral{
public:
enum PolyType { Even, Odd, Full };
private:
// In GSL-style, pass the function as a function pointer. Any data required to evaluate the function is passed in as a void pointer
bigfloat (*f)(bigfloat x, void *data);
void *data;
// The approximation parameters
std::vector<bigfloat> param;
bigfloat norm;
// The number of non-zero terms in the numerator and denominator
int n, d;
// The numerator and denominator degree (i.e. the largest power)
int pow_n, pow_d;
// Specify if the numerator and/or denominator are odd/even polynomials
PolyType num_type;
PolyType den_type;
std::vector<int> num_pows; //contains the mapping, with -1 if not present
std::vector<int> den_pows;
// The bounds of the approximation
bigfloat apstrt, apwidt, apend;
// Variables used to calculate the approximation
int nd1, iter;
std::vector<bigfloat> xx;
std::vector<bigfloat> mm;
std::vector<bigfloat> step;
bigfloat delta, spread;
// Variables used in search
std::vector<bigfloat> yy;
// Variables used in solving linear equations
std::vector<bigfloat> A;
std::vector<bigfloat> B;
std::vector<int> IPS;
// The number of equations we must solve at each iteration (n+d+1)
int neq;
// The precision of the GNU MP library
long prec;
// Initialize member variables associated with the polynomial's properties
void setupPolyProperties(int num_degree, int den_degree, PolyType num_type_in, PolyType den_type_in);
// Initial values of maximal and minmal errors
void initialGuess();
// Initialise step sizes
void stpini();
// Initialize the algorithm
void reinitializeAlgorithm();
// Solve the equations
void equations();
// Search for error maxima and minima
void search();
// Calculate function required for the approximation
inline bigfloat func(bigfloat x) const{
return f(x, data);
}
// Compute size and sign of the approximation error at x
bigfloat getErr(bigfloat x, int *sign) const;
// Solve the system AX=B where X = param
int simq();
// Evaluate the rational form P(x)/Q(x) using coefficients from the solution vector param
bigfloat approx(bigfloat x) const;
public:
AlgRemezGeneral(double lower, double upper, long prec,
bigfloat (*f)(bigfloat x, void *data), void *data);
inline int getDegree(void) const{
assert(n==d);
return n;
}
// Reset the bounds of the approximation
inline void setBounds(double lower, double upper) {
apstrt = lower;
apend = upper;
apwidt = apend - apstrt;
}
// Get the bounds of the approximation
inline void getBounds(double &lower, double &upper) const{
lower=(double)apstrt;
upper=(double)apend;
}
// Run the algorithm to generate the rational approximation
double generateApprox(int num_degree, int den_degree,
PolyType num_type, PolyType den_type,
const double tolerance = 1e-15, const int report_freq = 1000);
inline double generateApprox(int num_degree, int den_degree,
const double tolerance = 1e-15, const int report_freq = 1000){
return generateApprox(num_degree, den_degree, Full, Full, tolerance, report_freq);
}
// Evaluate the rational form P(x)/Q(x) using coefficients from the
// solution vector param
inline double evaluateApprox(double x) const{
return (double)approx((bigfloat)x);
}
// Evaluate the rational form Q(x)/P(x) using coefficients from the solution vector param
inline double evaluateInverseApprox(double x) const{
return 1.0/(double)approx((bigfloat)x);
}
// Calculate function required for the approximation
inline double evaluateFunc(double x) const{
return (double)func((bigfloat)x);
}
// Calculate inverse function required for the approximation
inline double evaluateInverseFunc(double x) const{
return 1.0/(double)func((bigfloat)x);
}
// Dump csv of function, approx and error
void csv(std::ostream &os = std::cout) const;
// Get the coefficient of the term x^i in the numerator
inline double getCoeffNum(const int i) const{
return num_pows[i] == -1 ? 0. : double(param[num_pows[i]]);
}
// Get the coefficient of the term x^i in the denominator
inline double getCoeffDen(const int i) const{
if(i == pow_d) return 1.0;
else return den_pows[i] == -1 ? 0. : double(param[den_pows[i]+n+1]);
}
};
#endif

183
Grid/algorithms/approx/ZMobius.cc Normal file
View File

@ -0,0 +1,183 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/algorithms/approx/ZMobius.cc
Copyright (C) 2015
Author: Christopher Kelly <ckelly@phys.columbia.edu>
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution directory
*************************************************************************************/
/* END LEGAL */
#include <Grid/algorithms/approx/ZMobius.h>
#include <Grid/algorithms/approx/RemezGeneral.h>
NAMESPACE_BEGIN(Grid);
NAMESPACE_BEGIN(Approx);
//Compute the tanh approximation
inline double epsilonMobius(const double x, const std::vector<ComplexD> &w){
int Ls = w.size();
ComplexD fxp = 1., fmp = 1.;
for(int i=0;i<Ls;i++){
fxp = fxp * ( w[i] + x );
fmp = fmp * ( w[i] - x );
}
return ((fxp - fmp)/(fxp + fmp)).real();
}
inline double epsilonMobius(const double x, const std::vector<RealD> &w){
int Ls = w.size();
double fxp = 1., fmp = 1.;
for(int i=0;i<Ls;i++){
fxp = fxp * ( w[i] + x );
fmp = fmp * ( w[i] - x );
}
return (fxp - fmp)/(fxp + fmp);
}
//Compute the tanh approximation in a form suitable for the Remez
bigfloat epsilonMobius(bigfloat x, void* data){
const std::vector<RealD> &omega = *( (std::vector<RealD> const*)data );
bigfloat fxp(1.0);
bigfloat fmp(1.0);
for(int i=0;i<omega.size();i++){
fxp = fxp * ( bigfloat(omega[i]) + x);
fmp = fmp * ( bigfloat(omega[i]) - x);
}
return (fxp - fmp)/(fxp + fmp);
}
//Compute the Zmobius Omega parameters suitable for eigenvalue range -lambda_bound <= lambda <= lambda_bound
//Note omega_i = 1/(b_i + c_i) where b_i and c_i are the Mobius parameters
void computeZmobiusOmega(std::vector<ComplexD> &omega_out, const int Ls_out,
const std::vector<RealD> &omega_in, const int Ls_in,
const RealD lambda_bound){
assert(omega_in.size() == Ls_in);
omega_out.resize(Ls_out);
//Use the Remez algorithm to generate the appropriate rational polynomial
//For odd polynomial, to satisfy Haar condition must take either positive or negative half of range (cf https://arxiv.org/pdf/0803.0439.pdf page 6)
AlgRemezGeneral remez(0, lambda_bound, 64, &epsilonMobius, (void*)&omega_in);
remez.generateApprox(Ls_out-1, Ls_out,AlgRemezGeneral::Odd, AlgRemezGeneral::Even, 1e-15, 100);
remez.csv(std::cout);
//The rational approximation has the form [ f(x) - f(-x) ] / [ f(x) + f(-x) ] where f(x) = \Prod_{i=0}^{L_s-1} ( \omega_i + x )
//cf https://academiccommons.columbia.edu/doi/10.7916/D8T72HD7 pg 102
//omega_i are therefore the negative of the complex roots of f(x)
//We can find the roots by recognizing that the eigenvalues of a matrix A are the roots of the characteristic polynomial
// \rho(\lambda) = det( A - \lambda I ) where I is the unit matrix
//The matrix whose characteristic polynomial is an arbitrary monic polynomial a0 + a1 x + a2 x^2 + ... x^n is the companion matrix
// A = | 0 1 0 0 0 .... 0 |
// | 0 0 1 0 0 .... 0 |
// | : : : : : : |
// | 0 0 0 0 0 1
// | -a0 -a1 -a2 ... ... -an|
//Note the Remez defines the largest power to have unit coefficient
std::vector<RealD> coeffs(Ls_out+1);
for(int i=0;i<Ls_out+1;i+=2) coeffs[i] = coeffs[i] = remez.getCoeffDen(i); //even powers
for(int i=1;i<Ls_out+1;i+=2) coeffs[i] = coeffs[i] = remez.getCoeffNum(i); //odd powers
std::vector<std::complex<RealD> > roots(Ls_out);
//Form the companion matrix
Eigen::MatrixXd compn(Ls_out,Ls_out);
for(int i=0;i<Ls_out-1;i++) compn(i,0) = 0.;
compn(Ls_out - 1, 0) = -coeffs[0];
for(int j=1;j<Ls_out;j++){
for(int i=0;i<Ls_out-1;i++) compn(i,j) = i == j-1 ? 1. : 0.;
compn(Ls_out - 1, j) = -coeffs[j];
}
//Eigensolve
Eigen::EigenSolver<Eigen::MatrixXd> slv(compn, false);
const auto & ev = slv.eigenvalues();
for(int i=0;i<Ls_out;i++)
omega_out[i] = -ev(i);
//Sort ascending (smallest at start of vector!)
std::sort(omega_out.begin(), omega_out.end(),
[&](const ComplexD &a, const ComplexD &b){ return a.real() < b.real() || (a.real() == b.real() && a.imag() < b.imag()); });
//McGlynn thesis pg 122 suggest improved iteration counts if magnitude of omega diminishes towards the center of the 5th dimension
std::vector<ComplexD> omega_tmp = omega_out;
int s_low=0, s_high=Ls_out-1, ss=0;
for(int s_from = Ls_out-1; s_from >= 0; s_from--){ //loop from largest omega
int s_to;
if(ss % 2 == 0){
s_to = s_low++;
}else{
s_to = s_high--;
}
omega_out[s_to] = omega_tmp[s_from];
++ss;
}
std::cout << "Resulting omega_i:" << std::endl;
for(int i=0;i<Ls_out;i++)
std::cout << omega_out[i] << std::endl;
std::cout << "Test result matches the approximate polynomial found by the Remez" << std::endl;
std::cout << "<x> <remez approx> <poly approx> <diff poly approx remez approx> <exact> <diff poly approx exact>\n";
int npt = 60;
double dlt = lambda_bound/double(npt-1);
for (int i =0; i<npt; i++){
double x = i*dlt;
double r = remez.evaluateApprox(x);
double p = epsilonMobius(x, omega_out);
double e = epsilonMobius(x, omega_in);
std::cout << x<< " " << r << " " << p <<" " <<r-p << " " << e << " " << e-p << std::endl;
}
}
//mobius_param = b+c with b-c=1
void computeZmobiusOmega(std::vector<ComplexD> &omega_out, const int Ls_out, const RealD mobius_param, const int Ls_in, const RealD lambda_bound){
std::vector<RealD> omega_in(Ls_in, 1./mobius_param);
computeZmobiusOmega(omega_out, Ls_out, omega_in, Ls_in, lambda_bound);
}
//ZMobius class takes gamma_i = (b+c) omega_i as its input, where b, c are factored out
void computeZmobiusGamma(std::vector<ComplexD> &gamma_out,
const RealD mobius_param_out, const int Ls_out,
const RealD mobius_param_in, const int Ls_in,
const RealD lambda_bound){
computeZmobiusOmega(gamma_out, Ls_out, mobius_param_in, Ls_in, lambda_bound);
for(int i=0;i<Ls_out;i++) gamma_out[i] = gamma_out[i] * mobius_param_out;
}
//Assumes mobius_param_out == mobius_param_in
void computeZmobiusGamma(std::vector<ComplexD> &gamma_out, const int Ls_out, const RealD mobius_param, const int Ls_in, const RealD lambda_bound){
computeZmobiusGamma(gamma_out, mobius_param, Ls_out, mobius_param, Ls_in, lambda_bound);
}
NAMESPACE_END(Approx);
NAMESPACE_END(Grid);

57
Grid/algorithms/approx/ZMobius.h Normal file
View File

@ -0,0 +1,57 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/algorithms/approx/ZMobius.h
Copyright (C) 2015
Author: Christopher Kelly <ckelly@phys.columbia.edu>
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution directory
*************************************************************************************/
/* END LEGAL */
#ifndef GRID_ZMOBIUS_APPROX_H
#define GRID_ZMOBIUS_APPROX_H
#include <Grid/GridCore.h>
NAMESPACE_BEGIN(Grid);
NAMESPACE_BEGIN(Approx);
//Compute the Zmobius Omega parameters suitable for eigenvalue range -lambda_bound <= lambda <= lambda_bound
//Note omega_i = 1/(b_i + c_i) where b_i and c_i are the Mobius parameters
void computeZmobiusOmega(std::vector<ComplexD> &omega_out, const int Ls_out,
const std::vector<RealD> &omega_in, const int Ls_in,
const RealD lambda_bound);
//mobius_param = b+c with b-c=1
void computeZmobiusOmega(std::vector<ComplexD> &omega_out, const int Ls_out, const RealD mobius_param, const int Ls_in, const RealD lambda_bound);
//ZMobius class takes gamma_i = (b+c) omega_i as its input, where b, c are factored out
void computeZmobiusGamma(std::vector<ComplexD> &gamma_out,
const RealD mobius_param_out, const int Ls_out,
const RealD mobius_param_in, const int Ls_in,
const RealD lambda_bound);
//Assumes mobius_param_out == mobius_param_in
void computeZmobiusGamma(std::vector<ComplexD> &gamma_out, const int Ls_out, const RealD mobius_param, const int Ls_in, const RealD lambda_bound);
NAMESPACE_END(Approx);
NAMESPACE_END(Grid);
#endif

6
Grid/algorithms/approx/bigfloat_double.h
View File

@ -25,6 +25,10 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
See the full license in the file "LICENSE" in the top level distribution directory
*************************************************************************************/
/* END LEGAL */
#ifndef INCLUDED_BIGFLOAT_DOUBLE_H
#define INCLUDED_BIGFLOAT_DOUBLE_H
#include <math.h>
typedef double mfloat;
@ -186,4 +190,6 @@ public:
// friend bigfloat& random(void);
};
#endif

222
Grid/algorithms/iterative/BiCGSTAB.h Normal file
View File

@ -0,0 +1,222 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/algorithms/iterative/BiCGSTAB.h
Copyright (C) 2015
Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk>
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
Author: paboyle <paboyle@ph.ed.ac.uk>
Author: juettner <juettner@soton.ac.uk>
Author: David Murphy <djmurphy@mit.edu>
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution
directory
*************************************************************************************/
/* END LEGAL */
#ifndef GRID_BICGSTAB_H
#define GRID_BICGSTAB_H
NAMESPACE_BEGIN(Grid);
/////////////////////////////////////////////////////////////
// Base classes for iterative processes based on operators
// single input vec, single output vec.
/////////////////////////////////////////////////////////////
template <class Field>
class BiCGSTAB : public OperatorFunction<Field>
{
public:
using OperatorFunction<Field>::operator();
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
BiCGSTAB(RealD tol, Integer maxit, bool err_on_no_conv = true) :
Tolerance(tol), MaxIterations(maxit), ErrorOnNoConverge(err_on_no_conv){};
void operator()(LinearOperatorBase<Field>& Linop, const Field& src, Field& psi)
{
psi.Checkerboard() = src.Checkerboard();
conformable(psi, src);
RealD cp(0), rho(1), rho_prev(0), alpha(1), beta(0), omega(1);
RealD a(0), bo(0), b(0), ssq(0);
Field p(src);
Field r(src);
Field rhat(src);
Field v(src);
Field s(src);
Field t(src);
Field h(src);
v = Zero();
p = Zero();
// Initial residual computation & set up
RealD guess = norm2(psi);
assert(std::isnan(guess) == 0);
Linop.Op(psi, v);
b = norm2(v);
r = src - v;
rhat = r;
a = norm2(r);
ssq = norm2(src);
std::cout << GridLogIterative << std::setprecision(8) << "BiCGSTAB: guess " << guess << std::endl;
std::cout << GridLogIterative << std::setprecision(8) << "BiCGSTAB: src " << ssq << std::endl;
std::cout << GridLogIterative << std::setprecision(8) << "BiCGSTAB: mp " << b << std::endl;
std::cout << GridLogIterative << std::setprecision(8) << "BiCGSTAB: r " << a << std::endl;
RealD rsq = Tolerance * Tolerance * ssq;
// Check if guess is really REALLY good :)
if(a <= rsq){ return; }
std::cout << GridLogIterative << std::setprecision(8) << "BiCGSTAB: k=0 residual " << a << " target " << rsq << std::endl;
GridStopWatch LinalgTimer;
GridStopWatch InnerTimer;
GridStopWatch AxpyNormTimer;
GridStopWatch LinearCombTimer;
GridStopWatch MatrixTimer;
GridStopWatch SolverTimer;
SolverTimer.Start();
int k;
for (k = 1; k <= MaxIterations; k++)
{
rho_prev = rho;
LinalgTimer.Start();
InnerTimer.Start();
ComplexD Crho = innerProduct(rhat,r);
InnerTimer.Stop();
rho = Crho.real();
beta = (rho / rho_prev) * (alpha / omega);
LinearCombTimer.Start();
bo = beta * omega;
auto p_v = p.View();
auto r_v = r.View();
auto v_v = v.View();
accelerator_for(ss, p_v.size(), Field::vector_object::Nsimd(),{
coalescedWrite(p_v[ss], beta*p_v(ss) - bo*v_v(ss) + r_v(ss));
});
LinearCombTimer.Stop();
LinalgTimer.Stop();
MatrixTimer.Start();
Linop.Op(p,v);
MatrixTimer.Stop();
LinalgTimer.Start();
InnerTimer.Start();
ComplexD Calpha = innerProduct(rhat,v);
InnerTimer.Stop();
alpha = rho / Calpha.real();
LinearCombTimer.Start();
auto h_v = h.View();
auto psi_v = psi.View();
accelerator_for(ss, h_v.size(), Field::vector_object::Nsimd(),{
coalescedWrite(h_v[ss], alpha*p_v(ss) + psi_v(ss));
});
auto s_v = s.View();
accelerator_for(ss, s_v.size(), Field::vector_object::Nsimd(),{
coalescedWrite(s_v[ss], -alpha*v_v(ss) + r_v(ss));
});
LinearCombTimer.Stop();
LinalgTimer.Stop();
MatrixTimer.Start();
Linop.Op(s,t);
MatrixTimer.Stop();
LinalgTimer.Start();
InnerTimer.Start();
ComplexD Comega = innerProduct(t,s);
InnerTimer.Stop();
omega = Comega.real() / norm2(t);
LinearCombTimer.Start();
auto t_v = t.View();
accelerator_for(ss, psi_v.size(), Field::vector_object::Nsimd(),{
coalescedWrite(psi_v[ss], h_v(ss) + omega * s_v(ss));
coalescedWrite(r_v[ss], -omega * t_v(ss) + s_v(ss));
});
LinearCombTimer.Stop();
cp = norm2(r);
LinalgTimer.Stop();
std::cout << GridLogIterative << "BiCGSTAB: Iteration " << k << " residual " << sqrt(cp/ssq) << " target " << Tolerance << std::endl;
// Stopping condition
if(cp <= rsq)
{
SolverTimer.Stop();
Linop.Op(psi, v);
p = v - src;
RealD srcnorm = sqrt(norm2(src));
RealD resnorm = sqrt(norm2(p));
RealD true_residual = resnorm / srcnorm;
std::cout << GridLogMessage << "BiCGSTAB 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;
std::cout << GridLogMessage << "\tInner " << InnerTimer.Elapsed() << std::endl;
std::cout << GridLogMessage << "\tAxpyNorm " << AxpyNormTimer.Elapsed() << std::endl;
std::cout << GridLogMessage << "\tLinearComb " << LinearCombTimer.Elapsed() << std::endl;
if(ErrorOnNoConverge){ assert(true_residual / Tolerance < 10000.0); }
IterationsToComplete = k;
return;
}
}
std::cout << GridLogMessage << "BiCGSTAB did NOT converge" << std::endl;
if(ErrorOnNoConverge){ assert(0); }
IterationsToComplete = k;
}
};
NAMESPACE_END(Grid);
#endif

158
Grid/algorithms/iterative/BiCGSTABMixedPrec.h Normal file
View File

@ -0,0 +1,158 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/algorithms/iterative/BiCGSTABMixedPrec.h
Copyright (C) 2015
Author: Christopher Kelly <ckelly@phys.columbia.edu>
Author: David Murphy <djmurphy@mit.edu>
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution directory
*************************************************************************************/
/* END LEGAL */
#ifndef GRID_BICGSTAB_MIXED_PREC_H
#define GRID_BICGSTAB_MIXED_PREC_H
NAMESPACE_BEGIN(Grid);
// Mixed precision restarted defect correction BiCGSTAB
template<class FieldD, class FieldF, typename std::enable_if< getPrecision<FieldD>::value == 2, int>::type = 0, typename std::enable_if< getPrecision<FieldF>::value == 1, int>::type = 0>
class MixedPrecisionBiCGSTAB : public LinearFunction<FieldD>
{
public:
RealD Tolerance;
RealD InnerTolerance; // Initial tolerance for inner CG. Defaults to Tolerance but can be changed
Integer MaxInnerIterations;
Integer MaxOuterIterations;
GridBase* SinglePrecGrid; // Grid for single-precision fields
RealD OuterLoopNormMult; // Stop the outer loop and move to a final double prec solve when the residual is OuterLoopNormMult * Tolerance
LinearOperatorBase<FieldF> &Linop_f;
LinearOperatorBase<FieldD> &Linop_d;
Integer TotalInnerIterations; //Number of inner CG iterations
Integer TotalOuterIterations; //Number of restarts
Integer TotalFinalStepIterations; //Number of CG iterations in final patch-up step
//Option to speed up *inner single precision* solves using a LinearFunction that produces a guess
LinearFunction<FieldF> *guesser;
MixedPrecisionBiCGSTAB(RealD tol, Integer maxinnerit, Integer maxouterit, GridBase* _sp_grid,
LinearOperatorBase<FieldF>& _Linop_f, LinearOperatorBase<FieldD>& _Linop_d) :
Linop_f(_Linop_f), Linop_d(_Linop_d), Tolerance(tol), InnerTolerance(tol), MaxInnerIterations(maxinnerit),
MaxOuterIterations(maxouterit), SinglePrecGrid(_sp_grid), OuterLoopNormMult(100.), guesser(NULL) {};
void useGuesser(LinearFunction<FieldF>& g){
guesser = &g;
}
void operator() (const FieldD& src_d_in, FieldD& sol_d)
{
TotalInnerIterations = 0;
GridStopWatch TotalTimer;
TotalTimer.Start();
int cb = src_d_in.Checkerboard();
sol_d.Checkerboard() = cb;
RealD src_norm = norm2(src_d_in);
RealD stop = src_norm * Tolerance*Tolerance;
GridBase* DoublePrecGrid = src_d_in.Grid();
FieldD tmp_d(DoublePrecGrid);
tmp_d.Checkerboard() = cb;
FieldD tmp2_d(DoublePrecGrid);
tmp2_d.Checkerboard() = cb;
FieldD src_d(DoublePrecGrid);
src_d = src_d_in; //source for next inner iteration, computed from residual during operation
RealD inner_tol = InnerTolerance;
FieldF src_f(SinglePrecGrid);
src_f.Checkerboard() = cb;
FieldF sol_f(SinglePrecGrid);
sol_f.Checkerboard() = cb;
BiCGSTAB<FieldF> CG_f(inner_tol, MaxInnerIterations);
CG_f.ErrorOnNoConverge = false;
GridStopWatch InnerCGtimer;
GridStopWatch PrecChangeTimer;
Integer &outer_iter = TotalOuterIterations; //so it will be equal to the final iteration count
for(outer_iter = 0; outer_iter < MaxOuterIterations; outer_iter++)
{
// Compute double precision rsd and also new RHS vector.
Linop_d.Op(sol_d, tmp_d);
RealD norm = axpy_norm(src_d, -1., tmp_d, src_d_in); //src_d is residual vector
std::cout << GridLogMessage << "MixedPrecisionBiCGSTAB: Outer iteration " << outer_iter << " residual " << norm << " target " << stop << std::endl;
if(norm < OuterLoopNormMult * stop){
std::cout << GridLogMessage << "MixedPrecisionBiCGSTAB: Outer iteration converged on iteration " << outer_iter << std::endl;
break;
}
while(norm * inner_tol * inner_tol < stop){ inner_tol *= 2; } // inner_tol = sqrt(stop/norm) ??
PrecChangeTimer.Start();
precisionChange(src_f, src_d);
PrecChangeTimer.Stop();
sol_f = Zero();
//Optionally improve inner solver guess (eg using known eigenvectors)
if(guesser != NULL){ (*guesser)(src_f, sol_f); }
//Inner CG
CG_f.Tolerance = inner_tol;
InnerCGtimer.Start();
CG_f(Linop_f, src_f, sol_f);
InnerCGtimer.Stop();
TotalInnerIterations += CG_f.IterationsToComplete;
//Convert sol back to double and add to double prec solution
PrecChangeTimer.Start();
precisionChange(tmp_d, sol_f);
PrecChangeTimer.Stop();
axpy(sol_d, 1.0, tmp_d, sol_d);
}
//Final trial CG
std::cout << GridLogMessage << "MixedPrecisionBiCGSTAB: Starting final patch-up double-precision solve" << std::endl;
BiCGSTAB<FieldD> CG_d(Tolerance, MaxInnerIterations);
CG_d(Linop_d, src_d_in, sol_d);
TotalFinalStepIterations = CG_d.IterationsToComplete;
TotalTimer.Stop();
std::cout << GridLogMessage << "MixedPrecisionBiCGSTAB: Inner CG iterations " << TotalInnerIterations << " Restarts " << TotalOuterIterations << " Final CG iterations " << TotalFinalStepIterations << std::endl;
std::cout << GridLogMessage << "MixedPrecisionBiCGSTAB: Total time " << TotalTimer.Elapsed() << " Precision change " << PrecChangeTimer.Elapsed() << " Inner CG total " << InnerCGtimer.Elapsed() << std::endl;
}
};
NAMESPACE_END(Grid);
#endif

12
Grid/algorithms/iterative/BlockConjugateGradient.h
View File

@ -52,6 +52,7 @@ class BlockConjugateGradient : public OperatorFunction<Field> {
Integer MaxIterations;
Integer IterationsToComplete; //Number of iterations the CG took to finish. Filled in upon completion
Integer PrintInterval; //GridLogMessages or Iterative
RealD TrueResidual;
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),PrintInterval(100)
@ -306,7 +307,8 @@ void BlockCGrQsolve(LinearOperatorBase<Field> &Linop, const Field &B, Field &X)
Linop.HermOp(X, AD);
AD = AD-B;
std::cout << GridLogMessage <<"\t True residual is " << std::sqrt(norm2(AD)/norm2(B)) <<std::endl;
TrueResidual = std::sqrt(norm2(AD)/norm2(B));
std::cout << GridLogMessage <<"\tTrue residual is " << TrueResidual <<std::endl;
std::cout << GridLogMessage << "Time Breakdown "<<std::endl;
std::cout << GridLogMessage << "\tElapsed " << SolverTimer.Elapsed() <<std::endl;
@ -442,7 +444,8 @@ void CGmultiRHSsolve(LinearOperatorBase<Field> &Linop, const Field &Src, Field &
Linop.HermOp(Psi, AP);
AP = AP-Src;
std::cout <<GridLogMessage << "\tTrue residual is " << std::sqrt(norm2(AP)/norm2(Src)) <<std::endl;
TrueResidual = std::sqrt(norm2(AP)/norm2(Src));
std::cout <<GridLogMessage << "\tTrue residual is " << TrueResidual <<std::endl;
std::cout << GridLogMessage << "Time Breakdown "<<std::endl;
std::cout << GridLogMessage << "\tElapsed " << SolverTimer.Elapsed() <<std::endl;
@ -653,7 +656,7 @@ void BlockCGrQsolveVec(LinearOperatorBase<Field> &Linop, const std::vector<Field
if ( rr > max_resid ) max_resid = rr;
}
std::cout << GridLogIterative << "\t Block Iteration "<<k<<" ave resid "<< sqrt(rrsum/sssum) << " max "<< sqrt(max_resid) <<std::endl;
std::cout << GridLogIterative << "\t Block Iteration "<<k<<" ave resid "<< std::sqrt(rrsum/sssum) << " max "<< std::sqrt(max_resid) <<std::endl;
if ( max_resid < Tolerance*Tolerance ) {
@ -668,7 +671,8 @@ void BlockCGrQsolveVec(LinearOperatorBase<Field> &Linop, const std::vector<Field
for(int b=0;b<Nblock;b++) Linop.HermOp(X[b], AD[b]);
for(int b=0;b<Nblock;b++) AD[b] = AD[b]-B[b];
std::cout << GridLogMessage <<"\t True residual is " << std::sqrt(normv(AD)/normv(B)) <<std::endl;
TrueResidual = std::sqrt(normv(AD)/normv(B));
std::cout << GridLogMessage << "\tTrue residual is " << TrueResidual <<std::endl;
std::cout << GridLogMessage << "Time Breakdown "<<std::endl;
std::cout << GridLogMessage << "\tElapsed " << SolverTimer.Elapsed() <<std::endl;

19
Grid/algorithms/iterative/ConjugateGradient.h
View File

@ -49,6 +49,7 @@ public:
RealD Tolerance;
Integer MaxIterations;
Integer IterationsToComplete; //Number of iterations the CG took to finish. Filled in upon completion
RealD TrueResidual;
ConjugateGradient(RealD tol, Integer maxit, bool err_on_no_conv = true)
: Tolerance(tol),
@ -81,6 +82,14 @@ public:
cp = a;
ssq = norm2(src);
// Handle trivial case of zero src
if (ssq == 0.){
psi = Zero();
IterationsToComplete = 1;
TrueResidual = 0.;
return;
}
std::cout << GridLogIterative << std::setprecision(8) << "ConjugateGradient: guess " << guess << std::endl;
std::cout << GridLogIterative << std::setprecision(8) << "ConjugateGradient: src " << ssq << std::endl;
std::cout << GridLogIterative << std::setprecision(8) << "ConjugateGradient: mp " << d << std::endl;
@ -92,6 +101,7 @@ public:
// Check if guess is really REALLY good :)
if (cp <= rsq) {
TrueResidual = std::sqrt(a/ssq);
std::cout << GridLogMessage << "ConjugateGradient guess is converged already " << std::endl;
IterationsToComplete = 0;
return;
@ -141,7 +151,7 @@ public:
LinalgTimer.Stop();
std::cout << GridLogIterative << "ConjugateGradient: Iteration " << k
<< " residual^2 " << sqrt(cp/ssq) << " target " << Tolerance << std::endl;
<< " residual " << sqrt(cp/ssq) << " target " << Tolerance << std::endl;
// Stopping condition
if (cp <= rsq) {
@ -169,10 +179,17 @@ public:
if (ErrorOnNoConverge) assert(true_residual / Tolerance < 10000.0);
IterationsToComplete = k;
TrueResidual = true_residual;
return;
}
}
// Failed. Calculate true residual before giving up
Linop.HermOpAndNorm(psi, mmp, d, qq);
p = mmp - src;
TrueResidual = sqrt(norm2(p)/ssq);
std::cout << GridLogMessage << "ConjugateGradient did NOT converge "<<k<<" / "<< MaxIterations<< std::endl;
if (ErrorOnNoConverge) assert(0);

21
Grid/algorithms/iterative/ConjugateGradientMultiShift.h
View File

@ -46,15 +46,19 @@ public:
RealD Tolerance;
Integer MaxIterations;
Integer IterationsToComplete; //Number of iterations the CG took to finish. Filled in upon completion
Integer IterationsToComplete; //Number of iterations the CG took to finish. Filled in upon completion
std::vector<int> IterationsToCompleteShift; // Iterations for this shift
int verbose;
MultiShiftFunction shifts;
std::vector<RealD> TrueResidualShift;
ConjugateGradientMultiShift(Integer maxit,MultiShiftFunction &_shifts) :
MaxIterations(maxit),
shifts(_shifts)
{
verbose=1;
IterationsToCompleteShift.resize(_shifts.order);
TrueResidualShift.resize(_shifts.order);
}
void operator() (LinearOperatorBase<Field> &Linop, const Field &src, Field &psi)
@ -125,6 +129,17 @@ public:
// Residuals "r" are src
// First search direction "p" is also src
cp = norm2(src);
// Handle trivial case of zero src.
if( cp == 0. ){
for(int s=0;s<nshift;s++){
psi[s] = Zero();
IterationsToCompleteShift[s] = 1;
TrueResidualShift[s] = 0.;
}
return;
}
for(int s=0;s<nshift;s++){
rsq[s] = cp * mresidual[s] * mresidual[s];
std::cout<<GridLogMessage<<"ConjugateGradientMultiShift: shift "<<s
@ -270,6 +285,7 @@ public:
for(int s=0;s<nshift;s++){
if ( (!converged[s]) ){
IterationsToCompleteShift[s] = k;
RealD css = c * z[s][iz]* z[s][iz];
@ -299,7 +315,8 @@ public:
axpy(r,-alpha[s],src,tmp);
RealD rn = norm2(r);
RealD cn = norm2(src);
std::cout<<GridLogMessage<<"CGMultiShift: shift["<<s<<"] true residual "<<std::sqrt(rn/cn)<<std::endl;
TrueResidualShift[s] = std::sqrt(rn/cn);
std::cout<<GridLogMessage<<"CGMultiShift: shift["<<s<<"] true residual "<< TrueResidualShift[s] <<std::endl;
}
std::cout << GridLogMessage << "Time Breakdown "<<std::endl;

135
Grid/algorithms/iterative/SchurRedBlack.h
View File

@ -405,6 +405,70 @@ namespace Grid {
}
};
template<class Field> class NonHermitianSchurRedBlackDiagMooeeSolve : public SchurRedBlackBase<Field>
{
public:
typedef CheckerBoardedSparseMatrixBase<Field> Matrix;
NonHermitianSchurRedBlackDiagMooeeSolve(OperatorFunction<Field>& RBSolver, const bool initSubGuess = false,
const bool _solnAsInitGuess = false)
: SchurRedBlackBase<Field>(RBSolver, initSubGuess, _solnAsInitGuess) {};
//////////////////////////////////////////////////////
// Override RedBlack specialisation
//////////////////////////////////////////////////////
virtual void RedBlackSource(Matrix& _Matrix, const Field& src, Field& src_e, Field& src_o)
{
GridBase* grid = _Matrix.RedBlackGrid();
GridBase* fgrid = _Matrix.Grid();
Field tmp(grid);
Field Mtmp(grid);
pickCheckerboard(Even, src_e, src);
pickCheckerboard(Odd , src_o, src);
/////////////////////////////////////////////////////
// src_o = Mdag * (source_o - Moe MeeInv source_e)
/////////////////////////////////////////////////////
_Matrix.MooeeInv(src_e, tmp); assert( tmp.Checkerboard() == Even );
_Matrix.Meooe (tmp, Mtmp); assert( Mtmp.Checkerboard() == Odd );
src_o -= Mtmp; assert( src_o.Checkerboard() == Odd );
}
virtual void RedBlackSolution(Matrix& _Matrix, const Field& sol_o, const Field& src_e, Field& sol)
{
GridBase* grid = _Matrix.RedBlackGrid();
GridBase* fgrid = _Matrix.Grid();
Field tmp(grid);
Field sol_e(grid);
Field src_e_i(grid);
///////////////////////////////////////////////////
// sol_e = M_ee^-1 * ( src_e - Meo sol_o )...
///////////////////////////////////////////////////
_Matrix.Meooe(sol_o, tmp); assert( tmp.Checkerboard() == Even );
src_e_i = src_e - tmp; assert( src_e_i.Checkerboard() == Even );
_Matrix.MooeeInv(src_e_i, sol_e); assert( sol_e.Checkerboard() == Even );
setCheckerboard(sol, sol_e); assert( sol_e.Checkerboard() == Even );
setCheckerboard(sol, sol_o); assert( sol_o.Checkerboard() == Odd );
}
virtual void RedBlackSolve(Matrix& _Matrix, const Field& src_o, Field& sol_o)
{
NonHermitianSchurDiagMooeeOperator<Matrix,Field> _OpEO(_Matrix);
this->_HermitianRBSolver(_OpEO, src_o, sol_o); assert(sol_o.Checkerboard() == Odd);
}
virtual void RedBlackSolve(Matrix& _Matrix, const std::vector<Field>& src_o, std::vector<Field>& sol_o)
{
NonHermitianSchurDiagMooeeOperator<Matrix,Field> _OpEO(_Matrix);
this->_HermitianRBSolver(_OpEO, src_o, sol_o);
}
};
///////////////////////////////////////////////////////////////////////////////////////////////////////
// Site diagonal is identity, right preconditioned by Mee^inv
// ( 1 - Meo Moo^inv Moe Mee^inv ) phi =( 1 - Meo Moo^inv Moe Mee^inv ) Mee psi = = eta = eta
@ -482,5 +546,76 @@ namespace Grid {
this->_HermitianRBSolver(_HermOpEO,src_o,sol_o);
}
};
template<class Field> class NonHermitianSchurRedBlackDiagTwoSolve : public SchurRedBlackBase<Field>
{
public:
typedef CheckerBoardedSparseMatrixBase<Field> Matrix;
/////////////////////////////////////////////////////
// Wrap the usual normal equations Schur trick
/////////////////////////////////////////////////////
NonHermitianSchurRedBlackDiagTwoSolve(OperatorFunction<Field>& RBSolver, const bool initSubGuess = false,
const bool _solnAsInitGuess = false)
: SchurRedBlackBase<Field>(RBSolver, initSubGuess, _solnAsInitGuess) {};
virtual void RedBlackSource(Matrix& _Matrix, const Field& src, Field& src_e, Field& src_o)
{
GridBase* grid = _Matrix.RedBlackGrid();
GridBase* fgrid = _Matrix.Grid();
Field tmp(grid);
Field Mtmp(grid);
pickCheckerboard(Even, src_e, src);
pickCheckerboard(Odd , src_o, src);
/////////////////////////////////////////////////////
// src_o = Mdag * (source_o - Moe MeeInv source_e)
/////////////////////////////////////////////////////
_Matrix.MooeeInv(src_e, tmp); assert( tmp.Checkerboard() == Even );
_Matrix.Meooe (tmp, Mtmp); assert( Mtmp.Checkerboard() == Odd );
src_o -= Mtmp; assert( src_o.Checkerboard() == Odd );
}
virtual void RedBlackSolution(Matrix& _Matrix, const Field& sol_o, const Field& src_e, Field& sol)
{
GridBase* grid = _Matrix.RedBlackGrid();
GridBase* fgrid = _Matrix.Grid();
Field sol_o_i(grid);
Field tmp(grid);
Field sol_e(grid);
////////////////////////////////////////////////
// MooeeInv due to pecond
////////////////////////////////////////////////
_Matrix.MooeeInv(sol_o, tmp);
sol_o_i = tmp;
///////////////////////////////////////////////////
// sol_e = M_ee^-1 * ( src_e - Meo sol_o )...
///////////////////////////////////////////////////
_Matrix.Meooe(sol_o_i, tmp); assert( tmp.Checkerboard() == Even );
tmp = src_e - tmp; assert( src_e.Checkerboard() == Even );
_Matrix.MooeeInv(tmp, sol_e); assert( sol_e.Checkerboard() == Even );
setCheckerboard(sol, sol_e); assert( sol_e.Checkerboard() == Even );
setCheckerboard(sol, sol_o_i); assert( sol_o_i.Checkerboard() == Odd );
};
virtual void RedBlackSolve(Matrix& _Matrix, const Field& src_o, Field& sol_o)
{
NonHermitianSchurDiagTwoOperator<Matrix,Field> _OpEO(_Matrix);
this->_HermitianRBSolver(_OpEO, src_o, sol_o);
};
virtual void RedBlackSolve(Matrix& _Matrix, const std::vector<Field>& src_o, std::vector<Field>& sol_o)
{
NonHermitianSchurDiagTwoOperator<Matrix,Field> _OpEO(_Matrix);
this->_HermitianRBSolver(_OpEO, src_o, sol_o);
}
};
}
#endif