portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-06-12 20:27:06 +01:00
Code Releases Activity

Merge remote-tracking branch 'grid/develop' into feature/arm-neon

Browse Source
This commit is contained in:
Nils Meyer
2017-08-29 17:47:36 +02:00
parent 7a53dc3715 4b4c2a715b
commit 4e907fef2c
165 changed files with 8491 additions and 5406 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
lib/Grid.h
View File

@ -41,7 +41,9 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
#include <Grid/GridCore.h>
#include <Grid/GridQCDcore.h>
#include <Grid/qcd/action/Action.h>
#include <Grid/qcd/utils/GaugeFix.h>
#include <Grid/qcd/smearing/Smearing.h>
#include <Grid/parallelIO/MetaData.h>
#include <Grid/qcd/hmc/HMC_aggregate.h>
#endif

2
lib/GridStd.h
View File

@ -7,6 +7,7 @@
#include <cassert>
#include <complex>
#include <vector>
#include <string>
#include <iostream>
#include <iomanip>
#include <random>
@ -18,6 +19,7 @@
#include <ctime>
#include <sys/time.h>
#include <chrono>
#include <zlib.h>
///////////////////
// Grid config

4
lib/Makefile.am
View File

@ -10,8 +10,8 @@ if BUILD_COMMS_MPI3
extra_sources+=communicator/Communicator_base.cc
endif
if BUILD_COMMS_MPI3L
extra_sources+=communicator/Communicator_mpi3_leader.cc
if BUILD_COMMS_MPIT
extra_sources+=communicator/Communicator_mpit.cc
extra_sources+=communicator/Communicator_base.cc
endif

137
lib/algorithms/densematrix/DenseMatrix.h
View File

@ -1,137 +0,0 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/algorithms/iterative/DenseMatrix.h
Copyright (C) 2015
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
Author: paboyle <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_DENSE_MATRIX_H
#define GRID_DENSE_MATRIX_H
namespace Grid {
/////////////////////////////////////////////////////////////
// Matrix untils
/////////////////////////////////////////////////////////////
template<class T> using DenseVector = std::vector<T>;
template<class T> using DenseMatrix = DenseVector<DenseVector<T> >;
template<class T> void Size(DenseVector<T> & vec, int &N)
{
N= vec.size();
}
template<class T> void Size(DenseMatrix<T> & mat, int &N,int &M)
{
N= mat.size();
M= mat[0].size();
}
template<class T> void SizeSquare(DenseMatrix<T> & mat, int &N)
{
int M; Size(mat,N,M);
assert(N==M);
}
template<class T> void Resize(DenseVector<T > & mat, int N) {
mat.resize(N);
}
template<class T> void Resize(DenseMatrix<T > & mat, int N, int M) {
mat.resize(N);
for(int i=0;i<N;i++){
mat[i].resize(M);
}
}
template<class T> void Fill(DenseMatrix<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 **/
template<class T> DenseMatrix<T> Transpose(DenseMatrix<T> & mat){
int N,M;
Size(mat,N,M);
DenseMatrix<T> 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 DenseMatrix to unit matrix **/
template<class T> void Unity(DenseMatrix<T> &A){
int N; SizeSquare(A,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(DenseMatrix<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 **/
template<class T>
DenseMatrix<T> HermitianConj(DenseMatrix<T> &mat){
int dim; SizeSquare(mat,dim);
DenseMatrix<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;
}
/**Get a square submatrix**/
template <class T>
DenseMatrix<T> GetSubMtx(DenseMatrix<T> &A,int row_st, int row_end, int col_st, int col_end)
{
DenseMatrix<T> H; Resize(H,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;
}
}
#include "Householder.h"
#include "Francis.h"
#endif

525
lib/algorithms/densematrix/Francis.h
View File

@ -1,525 +0,0 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/algorithms/iterative/Francis.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 FRANCIS_H
#define FRANCIS_H
#include <cstdlib>
#include <string>
#include <cmath>
#include <iostream>
#include <sstream>
#include <stdexcept>
#include <fstream>
#include <complex>
#include <algorithm>
//#include <timer.h>
//#include <lapacke.h>
//#include <Eigen/Dense>
namespace Grid {
template <class T> int SymmEigensystem(DenseMatrix<T > &Ain, DenseVector<T> &evals, DenseMatrix<T> &evecs, RealD small);
template <class T> int Eigensystem(DenseMatrix<T > &Ain, DenseVector<T> &evals, DenseMatrix<T> &evecs, RealD small);
/**
Find the eigenvalues of an upper hessenberg matrix using the Francis QR algorithm.
H =
x x x x x x x x x
x x x x x x x x x
0 x x x x x x x x
0 0 x x x x x x x
0 0 0 x x x x x x
0 0 0 0 x x x x x
0 0 0 0 0 x x x x
0 0 0 0 0 0 x x x
0 0 0 0 0 0 0 x x
Factorization is P T P^H where T is upper triangular (mod cc blocks) and P is orthagonal/unitary.
**/
template <class T>
int QReigensystem(DenseMatrix<T> &Hin, DenseVector<T> &evals, DenseMatrix<T> &evecs, RealD small)
{
DenseMatrix<T> H = Hin;
int N ; SizeSquare(H,N);
int M = N;
Fill(evals,0);
Fill(evecs,0);
T s,t,x=0,y=0,z=0;
T u,d;
T apd,amd,bc;
DenseVector<T> p(N,0);
T nrm = Norm(H); ///DenseMatrix Norm
int n, m;
int e = 0;
int it = 0;
int tot_it = 0;
int l = 0;
int r = 0;
DenseMatrix<T> P; Resize(P,N,N); Unity(P);
DenseVector<int> trows(N,0);
/// Check if the matrix is really hessenberg, if not abort
RealD sth = 0;
for(int j=0;j<N;j++){
for(int i=j+2;i<N;i++){
sth = abs(H[i][j]);
if(sth > small){
std::cout << "Non hessenberg H = " << sth << " > " << small << std::endl;
exit(1);
}
}
}
do{
std::cout << "Francis QR Step N = " << N << std::endl;
/** Check for convergence
x x x x x
0 x x x x
0 0 x x x
0 0 x x x
0 0 0 0 x
for this matrix l = 4
**/
do{
l = Chop_subdiag(H,nrm,e,small);
r = 0; ///May have converged on more than one eval
///Single eval
if(l == N-1){
evals[e] = H[l][l];
N--; e++; r++; it = 0;
}
///RealD eval
if(l == N-2){
trows[l+1] = 1; ///Needed for UTSolve
apd = H[l][l] + H[l+1][l+1];
amd = H[l][l] - H[l+1][l+1];
bc = (T)4.0*H[l+1][l]*H[l][l+1];
evals[e] = (T)0.5*( apd + sqrt(amd*amd + bc) );
evals[e+1] = (T)0.5*( apd - sqrt(amd*amd + bc) );
N-=2; e+=2; r++; it = 0;
}
} while(r>0);
if(N ==0) break;
DenseVector<T > ck; Resize(ck,3);
DenseVector<T> v; Resize(v,3);
for(int m = N-3; m >= l; m--){
///Starting vector essentially random shift.
if(it%10 == 0 && N >= 3 && it > 0){
s = (T)1.618033989*( abs( H[N-1][N-2] ) + abs( H[N-2][N-3] ) );
t = (T)0.618033989*( abs( H[N-1][N-2] ) + abs( H[N-2][N-3] ) );
x = H[m][m]*H[m][m] + H[m][m+1]*H[m+1][m] - s*H[m][m] + t;
y = H[m+1][m]*(H[m][m] + H[m+1][m+1] - s);
z = H[m+1][m]*H[m+2][m+1];
}
///Starting vector implicit Q theorem
else{
s = (H[N-2][N-2] + H[N-1][N-1]);
t = (H[N-2][N-2]*H[N-1][N-1] - H[N-2][N-1]*H[N-1][N-2]);
x = H[m][m]*H[m][m] + H[m][m+1]*H[m+1][m] - s*H[m][m] + t;
y = H[m+1][m]*(H[m][m] + H[m+1][m+1] - s);
z = H[m+1][m]*H[m+2][m+1];
}
ck[0] = x; ck[1] = y; ck[2] = z;
if(m == l) break;
/** Some stupid thing from numerical recipies, seems to work**/
// PAB.. for heaven's sake quote page, purpose, evidence it works.
// what sort of comment is that!?!?!?
u=abs(H[m][m-1])*(abs(y)+abs(z));
d=abs(x)*(abs(H[m-1][m-1])+abs(H[m][m])+abs(H[m+1][m+1]));
if ((T)abs(u+d) == (T)abs(d) ){
l = m; break;
}
//if (u < small){l = m; break;}
}
if(it > 100000){
std::cout << "QReigensystem: bugger it got stuck after 100000 iterations" << std::endl;
std::cout << "got " << e << " evals " << l << " " << N << std::endl;
exit(1);
}
normalize(ck); ///Normalization cancels in PHP anyway
T beta;
Householder_vector<T >(ck, 0, 2, v, beta);
Householder_mult<T >(H,v,beta,0,l,l+2,0);
Householder_mult<T >(H,v,beta,0,l,l+2,1);
///Accumulate eigenvector
Householder_mult<T >(P,v,beta,0,l,l+2,1);
int sw = 0; ///Are we on the last row?
for(int k=l;k<N-2;k++){
x = H[k+1][k];
y = H[k+2][k];
z = (T)0.0;
if(k+3 <= N-1){
z = H[k+3][k];
} else{
sw = 1;
v[2] = (T)0.0;
}
ck[0] = x; ck[1] = y; ck[2] = z;
normalize(ck);
Householder_vector<T >(ck, 0, 2-sw, v, beta);
Householder_mult<T >(H,v, beta,0,k+1,k+3-sw,0);
Householder_mult<T >(H,v, beta,0,k+1,k+3-sw,1);
///Accumulate eigenvector
Householder_mult<T >(P,v, beta,0,k+1,k+3-sw,1);
}
it++;
tot_it++;
}while(N > 1);
N = evals.size();
///Annoying - UT solves in reverse order;
DenseVector<T> tmp; Resize(tmp,N);
for(int i=0;i<N;i++){
tmp[i] = evals[N-i-1];
}
evals = tmp;
UTeigenvectors(H, trows, evals, evecs);
for(int i=0;i<evals.size();i++){evecs[i] = P*evecs[i]; normalize(evecs[i]);}
return tot_it;
}
template <class T>
int my_Wilkinson(DenseMatrix<T> &Hin, DenseVector<T> &evals, DenseMatrix<T> &evecs, RealD small)
{
/**
Find the eigenvalues of an upper Hessenberg matrix using the Wilkinson QR algorithm.
H =
x x 0 0 0 0
x x x 0 0 0
0 x x x 0 0
0 0 x x x 0
0 0 0 x x x
0 0 0 0 x x
Factorization is P T P^H where T is upper triangular (mod cc blocks) and P is orthagonal/unitary. **/
return my_Wilkinson(Hin, evals, evecs, small, small);
}
template <class T>
int my_Wilkinson(DenseMatrix<T> &Hin, DenseVector<T> &evals, DenseMatrix<T> &evecs, RealD small, RealD tol)
{
int N; SizeSquare(Hin,N);
int M = N;
///I don't want to modify the input but matricies must be passed by reference
//Scale a matrix by its "norm"
//RealD Hnorm = abs( Hin.LargestDiag() ); H = H*(1.0/Hnorm);
DenseMatrix<T> H; H = Hin;
RealD Hnorm = abs(Norm(Hin));
H = H * (1.0 / Hnorm);
// TODO use openmp and memset
Fill(evals,0);
Fill(evecs,0);
T s, t, x = 0, y = 0, z = 0;
T u, d;
T apd, amd, bc;
DenseVector<T> p; Resize(p,N); Fill(p,0);
T nrm = Norm(H); ///DenseMatrix Norm
int n, m;
int e = 0;
int it = 0;
int tot_it = 0;
int l = 0;
int r = 0;
DenseMatrix<T> P; Resize(P,N,N);
Unity(P);
DenseVector<int> trows(N, 0);
/// Check if the matrix is really symm tridiag
RealD sth = 0;
for(int j = 0; j < N; ++j)
{
for(int i = j + 2; i < N; ++i)
{
if(abs(H[i][j]) > tol || abs(H[j][i]) > tol)
{
std::cout << "Non Tridiagonal H(" << i << ","<< j << ") = |" << Real( real( H[j][i] ) ) << "| > " << tol << std::endl;
std::cout << "Warning tridiagonalize and call again" << std::endl;
// exit(1); // see what is going on
//return;
}
}
}
do{
do{
//Jasper
//Check if the subdiagonal term is small enough (<small)
//if true then it is converged.
//check start from H.dim - e - 1
//How to deal with more than 2 are converged?
//What if Chop_symm_subdiag return something int the middle?
//--------------
l = Chop_symm_subdiag(H,nrm, e, small);
r = 0; ///May have converged on more than one eval
//Jasper
//In this case
// x x 0 0 0 0
// x x x 0 0 0
// 0 x x x 0 0
// 0 0 x x x 0
// 0 0 0 x x 0
// 0 0 0 0 0 x <- l
//--------------
///Single eval
if(l == N - 1)
{
evals[e] = H[l][l];
N--;
e++;
r++;
it = 0;
}
//Jasper
// x x 0 0 0 0
// x x x 0 0 0
// 0 x x x 0 0
// 0 0 x x 0 0
// 0 0 0 0 x x <- l
// 0 0 0 0 x x
//--------------
///RealD eval
if(l == N - 2)
{
trows[l + 1] = 1; ///Needed for UTSolve
apd = H[l][l] + H[l + 1][ l + 1];
amd = H[l][l] - H[l + 1][l + 1];
bc = (T) 4.0 * H[l + 1][l] * H[l][l + 1];
evals[e] = (T) 0.5 * (apd + sqrt(amd * amd + bc));
evals[e + 1] = (T) 0.5 * (apd - sqrt(amd * amd + bc));
N -= 2;
e += 2;
r++;
it = 0;
}
}while(r > 0);
//Jasper
//Already converged
//--------------
if(N == 0) break;
DenseVector<T> ck,v; Resize(ck,2); Resize(v,2);
for(int m = N - 3; m >= l; m--)
{
///Starting vector essentially random shift.
if(it%10 == 0 && N >= 3 && it > 0)
{
t = abs(H[N - 1][N - 2]) + abs(H[N - 2][N - 3]);
x = H[m][m] - t;
z = H[m + 1][m];
} else {
///Starting vector implicit Q theorem
d = (H[N - 2][N - 2] - H[N - 1][N - 1]) * (T) 0.5;
t = H[N - 1][N - 1] - H[N - 1][N - 2] * H[N - 1][N - 2]
/ (d + sign(d) * sqrt(d * d + H[N - 1][N - 2] * H[N - 1][N - 2]));
x = H[m][m] - t;
z = H[m + 1][m];
}
//Jasper
//why it is here????
//-----------------------
if(m == l)
break;
u = abs(H[m][m - 1]) * (abs(y) + abs(z));
d = abs(x) * (abs(H[m - 1][m - 1]) + abs(H[m][m]) + abs(H[m + 1][m + 1]));
if ((T)abs(u + d) == (T)abs(d))
{
l = m;
break;
}
}
//Jasper
if(it > 1000000)
{
std::cout << "Wilkinson: bugger it got stuck after 100000 iterations" << std::endl;
std::cout << "got " << e << " evals " << l << " " << N << std::endl;
exit(1);
}
//
T s, c;
Givens_calc<T>(x, z, c, s);
Givens_mult<T>(H, l, l + 1, c, -s, 0);
Givens_mult<T>(H, l, l + 1, c, s, 1);
Givens_mult<T>(P, l, l + 1, c, s, 1);
//
for(int k = l; k < N - 2; ++k)
{
x = H.A[k + 1][k];
z = H.A[k + 2][k];
Givens_calc<T>(x, z, c, s);
Givens_mult<T>(H, k + 1, k + 2, c, -s, 0);
Givens_mult<T>(H, k + 1, k + 2, c, s, 1);
Givens_mult<T>(P, k + 1, k + 2, c, s, 1);
}
it++;
tot_it++;
}while(N > 1);
N = evals.size();
///Annoying - UT solves in reverse order;
DenseVector<T> tmp(N);
for(int i = 0; i < N; ++i)
tmp[i] = evals[N-i-1];
evals = tmp;
//
UTeigenvectors(H, trows, evals, evecs);
//UTSymmEigenvectors(H, trows, evals, evecs);
for(int i = 0; i < evals.size(); ++i)
{
evecs[i] = P * evecs[i];
normalize(evecs[i]);
evals[i] = evals[i] * Hnorm;
}
// // FIXME this is to test
// Hin.write("evecs3", evecs);
// Hin.write("evals3", evals);
// // check rsd
// for(int i = 0; i < M; i++) {
// vector<T> Aevec = Hin * evecs[i];
// RealD norm2(0.);
// for(int j = 0; j < M; j++) {
// norm2 += (Aevec[j] - evals[i] * evecs[i][j]) * (Aevec[j] - evals[i] * evecs[i][j]);
// }
// }
return tot_it;
}
template <class T>
void Hess(DenseMatrix<T > &A, DenseMatrix<T> &Q, int start){
/**
turn a matrix A =
x x x x x
x x x x x
x x x x x
x x x x x
x x x x x
into
x x x x x
x x x x x
0 x x x x
0 0 x x x
0 0 0 x x
with householder rotations
Slow.
*/
int N ; SizeSquare(A,N);
DenseVector<T > p; Resize(p,N); Fill(p,0);
for(int k=start;k<N-2;k++){
//cerr << "hess" << k << std::endl;
DenseVector<T > ck,v; Resize(ck,N-k-1); Resize(v,N-k-1);
for(int i=k+1;i<N;i++){ck[i-k-1] = A(i,k);} ///kth column
normalize(ck); ///Normalization cancels in PHP anyway
T beta;
Householder_vector<T >(ck, 0, ck.size()-1, v, beta); ///Householder vector
Householder_mult<T>(A,v,beta,start,k+1,N-1,0); ///A -> PA
Householder_mult<T >(A,v,beta,start,k+1,N-1,1); ///PA -> PAP^H
///Accumulate eigenvector
Householder_mult<T >(Q,v,beta,start,k+1,N-1,1); ///Q -> QP^H
}
/*for(int l=0;l<N-2;l++){
for(int k=l+2;k<N;k++){
A(0,k,l);
}
}*/
}
template <class T>
void Tri(DenseMatrix<T > &A, DenseMatrix<T> &Q, int start){
///Tridiagonalize a matrix
int N; SizeSquare(A,N);
Hess(A,Q,start);
/*for(int l=0;l<N-2;l++){
for(int k=l+2;k<N;k++){
A(0,l,k);
}
}*/
}
template <class T>
void ForceTridiagonal(DenseMatrix<T> &A){
///Tridiagonalize a matrix
int N ; SizeSquare(A,N);
for(int l=0;l<N-2;l++){
for(int k=l+2;k<N;k++){
A[l][k]=0;
A[k][l]=0;
}
}
}
template <class T>
int my_SymmEigensystem(DenseMatrix<T > &Ain, DenseVector<T> &evals, DenseVector<DenseVector<T> > &evecs, RealD small){
///Solve a symmetric eigensystem, not necessarily in tridiagonal form
int N; SizeSquare(Ain,N);
DenseMatrix<T > A; A = Ain;
DenseMatrix<T > Q; Resize(Q,N,N); Unity(Q);
Tri(A,Q,0);
int it = my_Wilkinson<T>(A, evals, evecs, small);
for(int k=0;k<N;k++){evecs[k] = Q*evecs[k];}
return it;
}
template <class T>
int Wilkinson(DenseMatrix<T> &Ain, DenseVector<T> &evals, DenseVector<DenseVector<T> > &evecs, RealD small){
return my_Wilkinson(Ain, evals, evecs, small);
}
template <class T>
int SymmEigensystem(DenseMatrix<T> &Ain, DenseVector<T> &evals, DenseVector<DenseVector<T> > &evecs, RealD small){
return my_SymmEigensystem(Ain, evals, evecs, small);
}
template <class T>
int Eigensystem(DenseMatrix<T > &Ain, DenseVector<T> &evals, DenseVector<DenseVector<T> > &evecs, RealD small){
///Solve a general eigensystem, not necessarily in tridiagonal form
int N = Ain.dim;
DenseMatrix<T > A(N); A = Ain;
DenseMatrix<T > Q(N);Q.Unity();
Hess(A,Q,0);
int it = QReigensystem<T>(A, evals, evecs, small);
for(int k=0;k<N;k++){evecs[k] = Q*evecs[k];}
return it;
}
}
#endif

242
lib/algorithms/densematrix/Householder.h
View File

@ -1,242 +0,0 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/algorithms/iterative/Householder.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 HOUSEHOLDER_H
#define HOUSEHOLDER_H
#define TIMER(A) std::cout << GridLogMessage << __FUNC__ << " file "<< __FILE__ <<" line " << __LINE__ << std::endl;
#define ENTER() std::cout << GridLogMessage << "ENTRY "<<__FUNC__ << " file "<< __FILE__ <<" line " << __LINE__ << std::endl;
#define LEAVE() std::cout << GridLogMessage << "EXIT "<<__FUNC__ << " file "<< __FILE__ <<" line " << __LINE__ << std::endl;
#include <cstdlib>
#include <string>
#include <cmath>
#include <iostream>
#include <sstream>
#include <stdexcept>
#include <fstream>
#include <complex>
#include <algorithm>
namespace Grid {
/** Comparison function for finding the max element in a vector **/
template <class T> bool cf(T i, T j) {
return abs(i) < abs(j);
}
/**
Calculate a real Givens angle
**/
template <class T> inline void Givens_calc(T y, T z, T &c, T &s){
RealD mz = (RealD)abs(z);
if(mz==0.0){
c = 1; s = 0;
}
if(mz >= (RealD)abs(y)){
T t = -y/z;
s = (T)1.0 / sqrt ((T)1.0 + t * t);
c = s * t;
} else {
T t = -z/y;
c = (T)1.0 / sqrt ((T)1.0 + t * t);
s = c * t;
}
}
template <class T> inline void Givens_mult(DenseMatrix<T> &A, int i, int k, T c, T s, int dir)
{
int q ; SizeSquare(A,q);
if(dir == 0){
for(int j=0;j<q;j++){
T nu = A[i][j];
T w = A[k][j];
A[i][j] = (c*nu + s*w);
A[k][j] = (-s*nu + c*w);
}
}
if(dir == 1){
for(int j=0;j<q;j++){
T nu = A[j][i];
T w = A[j][k];
A[j][i] = (c*nu - s*w);
A[j][k] = (s*nu + c*w);
}
}
}
/**
from input = x;
Compute the complex Householder vector, v, such that
P = (I - b v transpose(v) )
b = 2/v.v
P | x | | x | k = 0
| x | | 0 |
| x | = | 0 |
| x | | 0 | j = 3
| x | | x |
These are the "Unreduced" Householder vectors.
**/
template <class T> inline void Householder_vector(DenseVector<T> input, int k, int j, DenseVector<T> &v, T &beta)
{
int N ; Size(input,N);
T m = *max_element(input.begin() + k, input.begin() + j + 1, cf<T> );
if(abs(m) > 0.0){
T alpha = 0;
for(int i=k; i<j+1; i++){
v[i] = input[i]/m;
alpha = alpha + v[i]*conj(v[i]);
}
alpha = sqrt(alpha);
beta = (T)1.0/(alpha*(alpha + abs(v[k]) ));
if(abs(v[k]) > 0.0) v[k] = v[k] + (v[k]/abs(v[k]))*alpha;
else v[k] = -alpha;
} else{
for(int i=k; i<j+1; i++){
v[i] = 0.0;
}
}
}
/**
from input = x;
Compute the complex Householder vector, v, such that
P = (I - b v transpose(v) )
b = 2/v.v
Px = alpha*e_dir
These are the "Unreduced" Householder vectors.
**/
template <class T> inline void Householder_vector(DenseVector<T> input, int k, int j, int dir, DenseVector<T> &v, T &beta)
{
int N = input.size();
T m = *max_element(input.begin() + k, input.begin() + j + 1, cf);
if(abs(m) > 0.0){
T alpha = 0;
for(int i=k; i<j+1; i++){
v[i] = input[i]/m;
alpha = alpha + v[i]*conj(v[i]);
}
alpha = sqrt(alpha);
beta = 1.0/(alpha*(alpha + abs(v[dir]) ));
if(abs(v[dir]) > 0.0) v[dir] = v[dir] + (v[dir]/abs(v[dir]))*alpha;
else v[dir] = -alpha;
}else{
for(int i=k; i<j+1; i++){
v[i] = 0.0;
}
}
}
/**
Compute the product PA if trans = 0
AP if trans = 1
P = (I - b v transpose(v) )
b = 2/v.v
start at element l of matrix A
v is of length j - k + 1 of v are nonzero
**/
template <class T> inline void Householder_mult(DenseMatrix<T> &A , DenseVector<T> v, T beta, int l, int k, int j, int trans)
{
int N ; SizeSquare(A,N);
if(abs(beta) > 0.0){
for(int p=l; p<N; p++){
T s = 0;
if(trans==0){
for(int i=k;i<j+1;i++) s += conj(v[i-k])*A[i][p];
s *= beta;
for(int i=k;i<j+1;i++){ A[i][p] = A[i][p]-s*conj(v[i-k]);}
} else {
for(int i=k;i<j+1;i++){ s += conj(v[i-k])*A[p][i];}
s *= beta;
for(int i=k;i<j+1;i++){ A[p][i]=A[p][i]-s*conj(v[i-k]);}
}
}
}
}
/**
Compute the product PA if trans = 0
AP if trans = 1
P = (I - b v transpose(v) )
b = 2/v.v
start at element l of matrix A
v is of length j - k + 1 of v are nonzero
A is tridiagonal
**/
template <class T> inline void Householder_mult_tri(DenseMatrix<T> &A , DenseVector<T> v, T beta, int l, int M, int k, int j, int trans)
{
if(abs(beta) > 0.0){
int N ; SizeSquare(A,N);
DenseMatrix<T> tmp; Resize(tmp,N,N); Fill(tmp,0);
T s;
for(int p=l; p<M; p++){
s = 0;
if(trans==0){
for(int i=k;i<j+1;i++) s = s + conj(v[i-k])*A[i][p];
}else{
for(int i=k;i<j+1;i++) s = s + v[i-k]*A[p][i];
}
s = beta*s;
if(trans==0){
for(int i=k;i<j+1;i++) tmp[i][p] = tmp(i,p) - s*v[i-k];
}else{
for(int i=k;i<j+1;i++) tmp[p][i] = tmp[p][i] - s*conj(v[i-k]);
}
}
for(int p=l; p<M; p++){
if(trans==0){
for(int i=k;i<j+1;i++) A[i][p] = A[i][p] + tmp[i][p];
}else{
for(int i=k;i<j+1;i++) A[p][i] = A[p][i] + tmp[p][i];
}
}
}
}
}
#endif

314
lib/algorithms/iterative/BlockConjugateGradient.h
View File

@ -33,6 +33,8 @@ directory
namespace Grid {
enum BlockCGtype { BlockCG, BlockCGrQ, CGmultiRHS };
//////////////////////////////////////////////////////////////////////////
// Block conjugate gradient. Dimension zero should be the block direction
//////////////////////////////////////////////////////////////////////////
@ -40,25 +42,280 @@ template <class Field>
class BlockConjugateGradient : public OperatorFunction<Field> {
public:
typedef typename Field::scalar_type scomplex;
const int blockDim = 0;
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(RealD tol, Integer maxit, bool err_on_no_conv = true)
: Tolerance(tol),
MaxIterations(maxit),
ErrorOnNoConverge(err_on_no_conv){};
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)
{
int Orthog = 0; // First dimension is block dim
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;
//std::cout << GridLogMessage << " initial tmp " << norm2(tmp)<< std::endl;
ThinQRfact (m_rr, m_C, m_Cinv, Q, tmp);
//std::cout << GridLogMessage << " initial Q " << norm2(Q)<< std::endl;
//std::cout << GridLogMessage << " m_rr " << m_rr<<std::endl;
//std::cout << GridLogMessage << " m_C " << m_C<<std::endl;
//std::cout << GridLogMessage << " m_Cinv " << m_Cinv<<std::endl;
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();
//std::cout << GridLogMessage << " norm2 Z " <<norm2(Z)<<std::endl;
//4. M = [D^dag Z]^{-1}
sliceInnerTimer.Start();
sliceInnerProductMatrix(m_DZ,D,Z,Orthog);
sliceInnerTimer.Stop();
m_M = m_DZ.inverse();
//std::cout << GridLogMessage << " m_DZ " <<m_DZ<<std::endl;
//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];
std::cout<<GridLogMessage<<" Block Conjugate Gradient : Orthog "<<Orthog<<" Nblock "<<Nblock<<std::endl;
@ -162,8 +419,9 @@ 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;
}
@ -173,13 +431,14 @@ 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 <<"\tTrue 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;
@ -197,35 +456,13 @@ 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;
const int blockDim = 0;
int Nblock;
bool ErrorOnNoConverge; // throw an assert when the CG fails to converge.
// Defaults true.
RealD Tolerance;
Integer MaxIterations;
Integer IterationsToComplete; //Number of iterations the CG took to finish. Filled in upon completion
MultiRHSConjugateGradient(RealD tol, Integer maxit, bool err_on_no_conv = true)
: Tolerance(tol),
MaxIterations(maxit),
ErrorOnNoConverge(err_on_no_conv){};
void operator()(LinearOperatorBase<Field> &Linop, const Field &Src, Field &Psi)
void CGmultiRHSsolve(LinearOperatorBase<Field> &Linop, const Field &Src, Field &Psi)
{
int Orthog = 0; // First dimension is block dim
int Orthog = blockDim; // First dimension is block dim
Nblock = Src._grid->_fdimensions[Orthog];
std::cout<<GridLogMessage<<"MultiRHS Conjugate Gradient : Orthog "<<Orthog<<" Nblock "<<Nblock<<std::endl;
@ -285,12 +522,10 @@ void operator()(LinearOperatorBase<Field> &Linop, const Field &Src, Field &Psi)
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]);
}
@ -332,7 +567,7 @@ 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;
@ -358,9 +593,8 @@ void operator()(LinearOperatorBase<Field> &Linop, const Field &Src, Field &Psi)
if (ErrorOnNoConverge) assert(0);
IterationsToComplete = k;
}
};
}
#endif

81
lib/algorithms/iterative/EigenSort.h
View File

@ -1,81 +0,0 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/algorithms/iterative/EigenSort.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_EIGENSORT_H
#define GRID_EIGENSORT_H
namespace Grid {
/////////////////////////////////////////////////////////////
// Eigen sorter to begin with
/////////////////////////////////////////////////////////////
template<class Field>
class SortEigen {
private:
//hacking for testing for now
private:
static bool less_lmd(RealD left,RealD right){
return left > right;
}
static bool less_pair(std::pair<RealD,Field const*>& left,
std::pair<RealD,Field const*>& right){
return left.first > (right.first);
}
public:
void push(DenseVector<RealD>& lmd,
DenseVector<Field>& evec,int N) {
DenseVector<Field> cpy(lmd.size(),evec[0]._grid);
for(int i=0;i<lmd.size();i++) cpy[i] = evec[i];
DenseVector<std::pair<RealD, Field const*> > emod(lmd.size());
for(int i=0;i<lmd.size();++i)
emod[i] = std::pair<RealD,Field const*>(lmd[i],&cpy[i]);
partial_sort(emod.begin(),emod.begin()+N,emod.end(),less_pair);
typename DenseVector<std::pair<RealD, Field const*> >::iterator it = emod.begin();
for(int i=0;i<N;++i){
lmd[i]=it->first;
evec[i]=*(it->second);
++it;
}
}
void push(DenseVector<RealD>& lmd,int N) {
std::partial_sort(lmd.begin(),lmd.begin()+N,lmd.end(),less_lmd);
}
bool saturated(RealD lmd, RealD thrs) {
return fabs(lmd) > fabs(thrs);
}
};
}
#endif

1635
lib/algorithms/iterative/ImplicitlyRestartedLanczos.h
View File

File diff suppressed because it is too large Load Diff

39
lib/allocator/AlignedAllocator.cc
View File

@ -1,7 +1,5 @@
#include <Grid/GridCore.h>
#include <fcntl.h>
namespace Grid {
@ -11,7 +9,7 @@ int PointerCache::victim;
void *PointerCache::Insert(void *ptr,size_t bytes) {
if (bytes < 4096 ) return NULL;
if (bytes < 4096 ) return ptr;
#ifdef GRID_OMP
assert(omp_in_parallel()==0);
@ -63,4 +61,37 @@ void *PointerCache::Lookup(size_t bytes) {
return NULL;
}
void check_huge_pages(void *Buf,uint64_t BYTES)
{
#ifdef __linux__
int fd = open("/proc/self/pagemap", O_RDONLY);
assert(fd >= 0);
const int page_size = 4096;
uint64_t virt_pfn = (uint64_t)Buf / page_size;
off_t offset = sizeof(uint64_t) * virt_pfn;
uint64_t npages = (BYTES + page_size-1) / page_size;
uint64_t pagedata[npages];
uint64_t ret = lseek(fd, offset, SEEK_SET);
assert(ret == offset);
ret = ::read(fd, pagedata, sizeof(uint64_t)*npages);
assert(ret == sizeof(uint64_t) * npages);
int nhugepages = npages / 512;
int n4ktotal, nnothuge;
n4ktotal = 0;
nnothuge = 0;
for (int i = 0; i < nhugepages; ++i) {
uint64_t baseaddr = (pagedata[i*512] & 0x7fffffffffffffULL) * page_size;
for (int j = 0; j < 512; ++j) {
uint64_t pageaddr = (pagedata[i*512+j] & 0x7fffffffffffffULL) * page_size;
++n4ktotal;
if (pageaddr != baseaddr + j * page_size)
++nnothuge;
}
}
int rank = CartesianCommunicator::RankWorld();
printf("rank %d Allocated %d 4k pages, %d not in huge pages\n", rank, n4ktotal, nnothuge);
#endif
}
}

41
lib/allocator/AlignedAllocator.h
View File

@ -64,6 +64,8 @@ namespace Grid {
};
void check_huge_pages(void *Buf,uint64_t BYTES);
////////////////////////////////////////////////////////////////////
// A lattice of something, but assume the something is SIMDized.
////////////////////////////////////////////////////////////////////
@ -92,18 +94,34 @@ public:
size_type bytes = __n*sizeof(_Tp);
_Tp *ptr = (_Tp *) PointerCache::Lookup(bytes);
#ifdef HAVE_MM_MALLOC_H
if ( ptr == (_Tp *) NULL ) ptr = (_Tp *) _mm_malloc(bytes,128);
#else
if ( ptr == (_Tp *) NULL ) ptr = (_Tp *) memalign(128,bytes);
#endif
// if ( ptr != NULL )
// std::cout << "alignedAllocator "<<__n << " cache hit "<< std::hex << ptr <<std::dec <<std::endl;
//////////////////
// Hack 2MB align; could make option probably doesn't need configurability
//////////////////
//define GRID_ALLOC_ALIGN (128)
#define GRID_ALLOC_ALIGN (2*1024*1024)
#ifdef HAVE_MM_MALLOC_H
if ( ptr == (_Tp *) NULL ) ptr = (_Tp *) _mm_malloc(bytes,GRID_ALLOC_ALIGN);
#else
if ( ptr == (_Tp *) NULL ) ptr = (_Tp *) memalign(GRID_ALLOC_ALIGN,bytes);
#endif
// std::cout << "alignedAllocator " << std::hex << ptr <<std::dec <<std::endl;
// First touch optimise in threaded loop
uint8_t *cp = (uint8_t *)ptr;
#ifdef GRID_OMP
#pragma omp parallel for
#endif
for(size_type n=0;n<bytes;n+=4096){
cp[n]=0;
}
return ptr;
}
void deallocate(pointer __p, size_type __n) {
size_type bytes = __n * sizeof(_Tp);
pointer __freeme = (pointer)PointerCache::Insert((void *)__p,bytes);
#ifdef HAVE_MM_MALLOC_H
@ -182,10 +200,17 @@ public:
pointer allocate(size_type __n, const void* _p= 0)
{
#ifdef HAVE_MM_MALLOC_H
_Tp * ptr = (_Tp *) _mm_malloc(__n*sizeof(_Tp),128);
_Tp * ptr = (_Tp *) _mm_malloc(__n*sizeof(_Tp),GRID_ALLOC_ALIGN);
#else
_Tp * ptr = (_Tp *) memalign(128,__n*sizeof(_Tp));
_Tp * ptr = (_Tp *) memalign(GRID_ALLOC_ALIGN,__n*sizeof(_Tp));
#endif
size_type bytes = __n*sizeof(_Tp);
uint8_t *cp = (uint8_t *)ptr;
// One touch per 4k page, static OMP loop to catch same loop order
#pragma omp parallel for schedule(static)
for(size_type n=0;n<bytes;n+=4096){
cp[n]=0;
}
return ptr;
}
void deallocate(pointer __p, size_type) {

32
lib/cartesian/Cartesian_base.h
View File

@ -50,7 +50,6 @@ public:
GridBase(const std::vector<int> & processor_grid) : CartesianCommunicator(processor_grid) {};
// Physics Grid information.
std::vector<int> _simd_layout;// Which dimensions get relayed out over simd lanes.
std::vector<int> _fdimensions;// (full) Global dimensions of array prior to cb removal
@ -63,13 +62,12 @@ public:
int _isites;
int _fsites; // _isites*_osites = product(dimensions).
int _gsites;
std::vector<int> _slice_block; // subslice information
std::vector<int> _slice_block;// subslice information
std::vector<int> _slice_stride;
std::vector<int> _slice_nblock;
// Might need these at some point
// std::vector<int> _lstart; // local start of array in gcoors. _processor_coor[d]*_ldimensions[d]
// std::vector<int> _lend; // local end of array in gcoors _processor_coor[d]*_ldimensions[d]+_ldimensions_[d]-1
std::vector<int> _lstart; // local start of array in gcoors _processor_coor[d]*_ldimensions[d]
std::vector<int> _lend ; // local end of array in gcoors _processor_coor[d]*_ldimensions[d]+_ldimensions_[d]-1
public:
@ -176,6 +174,7 @@ public:
inline int gSites(void) const { return _isites*_osites*_Nprocessors; };
inline int Nd (void) const { return _ndimension;};
inline const std::vector<int> LocalStarts(void) { return _lstart; };
inline const std::vector<int> &FullDimensions(void) { return _fdimensions;};
inline const std::vector<int> &GlobalDimensions(void) { return _gdimensions;};
inline const std::vector<int> &LocalDimensions(void) { return _ldimensions;};
@ -186,17 +185,18 @@ public:
////////////////////////////////////////////////////////////////
void show_decomposition(){
std::cout << GridLogMessage << "Full Dimensions : " << _fdimensions << std::endl;
std::cout << GridLogMessage << "Global Dimensions : " << _gdimensions << std::endl;
std::cout << GridLogMessage << "Local Dimensions : " << _ldimensions << std::endl;
std::cout << GridLogMessage << "Reduced Dimensions : " << _rdimensions << std::endl;
std::cout << GridLogMessage << "Outer strides : " << _ostride << std::endl;
std::cout << GridLogMessage << "Inner strides : " << _istride << std::endl;
std::cout << GridLogMessage << "iSites : " << _isites << std::endl;
std::cout << GridLogMessage << "oSites : " << _osites << std::endl;
std::cout << GridLogMessage << "lSites : " << lSites() << std::endl;
std::cout << GridLogMessage << "gSites : " << gSites() << std::endl;
std::cout << GridLogMessage << "Nd : " << _ndimension << std::endl;
std::cout << GridLogMessage << "\tFull Dimensions : " << _fdimensions << std::endl;
std::cout << GridLogMessage << "\tSIMD layout : " << _simd_layout << std::endl;
std::cout << GridLogMessage << "\tGlobal Dimensions : " << _gdimensions << std::endl;
std::cout << GridLogMessage << "\tLocal Dimensions : " << _ldimensions << std::endl;
std::cout << GridLogMessage << "\tReduced Dimensions : " << _rdimensions << std::endl;
std::cout << GridLogMessage << "\tOuter strides : " << _ostride << std::endl;
std::cout << GridLogMessage << "\tInner strides : " << _istride << std::endl;
std::cout << GridLogMessage << "\tiSites : " << _isites << std::endl;
std::cout << GridLogMessage << "\toSites : " << _osites << std::endl;
std::cout << GridLogMessage << "\tlSites : " << lSites() << std::endl;
std::cout << GridLogMessage << "\tgSites : " << gSites() << std::endl;
std::cout << GridLogMessage << "\tNd : " << _ndimension << std::endl;
}
////////////////////////////////////////////////////////////////

126
lib/cartesian/Cartesian_full.h
View File

@ -62,73 +62,81 @@ public:
return shift;
}
GridCartesian(const std::vector<int> &dimensions,
const std::vector<int> &simd_layout,
const std::vector<int> &processor_grid
) : GridBase(processor_grid)
const std::vector<int> &simd_layout,
const std::vector<int> &processor_grid) : GridBase(processor_grid)
{
///////////////////////
// Grid information
///////////////////////
_ndimension = dimensions.size();
_fdimensions.resize(_ndimension);
_gdimensions.resize(_ndimension);
_ldimensions.resize(_ndimension);
_rdimensions.resize(_ndimension);
_simd_layout.resize(_ndimension);
_ostride.resize(_ndimension);
_istride.resize(_ndimension);
_fsites = _gsites = _osites = _isites = 1;
///////////////////////
// Grid information
///////////////////////
_ndimension = dimensions.size();
for(int d=0;d<_ndimension;d++){
_fdimensions[d] = dimensions[d]; // Global dimensions
_gdimensions[d] = _fdimensions[d]; // Global dimensions
_simd_layout[d] = simd_layout[d];
_fsites = _fsites * _fdimensions[d];
_gsites = _gsites * _gdimensions[d];
_fdimensions.resize(_ndimension);
_gdimensions.resize(_ndimension);
_ldimensions.resize(_ndimension);
_rdimensions.resize(_ndimension);
_simd_layout.resize(_ndimension);
_lstart.resize(_ndimension);
_lend.resize(_ndimension);
//FIXME check for exact division
_ostride.resize(_ndimension);
_istride.resize(_ndimension);
// Use a reduced simd grid
_ldimensions[d]= _gdimensions[d]/_processors[d]; //local dimensions
_rdimensions[d]= _ldimensions[d]/_simd_layout[d]; //overdecomposition
_osites *= _rdimensions[d];
_isites *= _simd_layout[d];
// Addressing support
if ( d==0 ) {
_ostride[d] = 1;
_istride[d] = 1;
} else {
_ostride[d] = _ostride[d-1]*_rdimensions[d-1];
_istride[d] = _istride[d-1]*_simd_layout[d-1];
}
_fsites = _gsites = _osites = _isites = 1;
for (int d = 0; d < _ndimension; d++)
{
_fdimensions[d] = dimensions[d]; // Global dimensions
_gdimensions[d] = _fdimensions[d]; // Global dimensions
_simd_layout[d] = simd_layout[d];
_fsites = _fsites * _fdimensions[d];
_gsites = _gsites * _gdimensions[d];
// Use a reduced simd grid
_ldimensions[d] = _gdimensions[d] / _processors[d]; //local dimensions
assert(_ldimensions[d] * _processors[d] == _gdimensions[d]);
_rdimensions[d] = _ldimensions[d] / _simd_layout[d]; //overdecomposition
assert(_rdimensions[d] * _simd_layout[d] == _ldimensions[d]);
_lstart[d] = _processor_coor[d] * _ldimensions[d];
_lend[d] = _processor_coor[d] * _ldimensions[d] + _ldimensions[d] - 1;
_osites *= _rdimensions[d];
_isites *= _simd_layout[d];
// Addressing support
if (d == 0)
{
_ostride[d] = 1;
_istride[d] = 1;
}
///////////////////////
// subplane information
///////////////////////
_slice_block.resize(_ndimension);
_slice_stride.resize(_ndimension);
_slice_nblock.resize(_ndimension);
int block =1;
int nblock=1;
for(int d=0;d<_ndimension;d++) nblock*=_rdimensions[d];
for(int d=0;d<_ndimension;d++){
nblock/=_rdimensions[d];
_slice_block[d] =block;
_slice_stride[d]=_ostride[d]*_rdimensions[d];
_slice_nblock[d]=nblock;
block = block*_rdimensions[d];
else
{
_ostride[d] = _ostride[d - 1] * _rdimensions[d - 1];
_istride[d] = _istride[d - 1] * _simd_layout[d - 1];
}
}
///////////////////////
// subplane information
///////////////////////
_slice_block.resize(_ndimension);
_slice_stride.resize(_ndimension);
_slice_nblock.resize(_ndimension);
int block = 1;
int nblock = 1;
for (int d = 0; d < _ndimension; d++)
nblock *= _rdimensions[d];
for (int d = 0; d < _ndimension; d++)
{
nblock /= _rdimensions[d];
_slice_block[d] = block;
_slice_stride[d] = _ostride[d] * _rdimensions[d];
_slice_nblock[d] = nblock;
block = block * _rdimensions[d];
}
};
};
}
#endif

191
lib/cartesian/Cartesian_red_black.h
View File

@ -131,132 +131,155 @@ public:
Init(dimensions,simd_layout,processor_grid,checker_dim_mask,0);
}
void Init(const std::vector<int> &dimensions,
const std::vector<int> &simd_layout,
const std::vector<int> &processor_grid,
const std::vector<int> &checker_dim_mask,
int checker_dim)
const std::vector<int> &simd_layout,
const std::vector<int> &processor_grid,
const std::vector<int> &checker_dim_mask,
int checker_dim)
{
///////////////////////
// Grid information
///////////////////////
///////////////////////
// Grid information
///////////////////////
_checker_dim = checker_dim;
assert(checker_dim_mask[checker_dim]==1);
assert(checker_dim_mask[checker_dim] == 1);
_ndimension = dimensions.size();
assert(checker_dim_mask.size()==_ndimension);
assert(processor_grid.size()==_ndimension);
assert(simd_layout.size()==_ndimension);
assert(checker_dim_mask.size() == _ndimension);
assert(processor_grid.size() == _ndimension);
assert(simd_layout.size() == _ndimension);
_fdimensions.resize(_ndimension);
_gdimensions.resize(_ndimension);
_ldimensions.resize(_ndimension);
_rdimensions.resize(_ndimension);
_simd_layout.resize(_ndimension);
_lstart.resize(_ndimension);
_lend.resize(_ndimension);
_ostride.resize(_ndimension);
_istride.resize(_ndimension);
_fsites = _gsites = _osites = _isites = 1;
_checker_dim_mask=checker_dim_mask;
for(int d=0;d<_ndimension;d++){
_fdimensions[d] = dimensions[d];
_gdimensions[d] = _fdimensions[d];
_fsites = _fsites * _fdimensions[d];
_gsites = _gsites * _gdimensions[d];
if (d==_checker_dim) {
_gdimensions[d] = _gdimensions[d]/2; // Remove a checkerboard
}
_ldimensions[d] = _gdimensions[d]/_processors[d];
_checker_dim_mask = checker_dim_mask;
// Use a reduced simd grid
_simd_layout[d] = simd_layout[d];
_rdimensions[d]= _ldimensions[d]/_simd_layout[d];
assert(_rdimensions[d]>0);
for (int d = 0; d < _ndimension; d++)
{
_fdimensions[d] = dimensions[d];
_gdimensions[d] = _fdimensions[d];
_fsites = _fsites * _fdimensions[d];
_gsites = _gsites * _gdimensions[d];
// all elements of a simd vector must have same checkerboard.
// If Ls vectorised, this must still be the case; e.g. dwf rb5d
if ( _simd_layout[d]>1 ) {
if ( checker_dim_mask[d] ) {
assert( (_rdimensions[d]&0x1) == 0 );
}
}
if (d == _checker_dim)
{
assert((_gdimensions[d] & 0x1) == 0);
_gdimensions[d] = _gdimensions[d] / 2; // Remove a checkerboard
}
_ldimensions[d] = _gdimensions[d] / _processors[d];
assert(_ldimensions[d] * _processors[d] == _gdimensions[d]);
_lstart[d] = _processor_coor[d] * _ldimensions[d];
_lend[d] = _processor_coor[d] * _ldimensions[d] + _ldimensions[d] - 1;
_osites *= _rdimensions[d];
_isites *= _simd_layout[d];
// Addressing support
if ( d==0 ) {
_ostride[d] = 1;
_istride[d] = 1;
} else {
_ostride[d] = _ostride[d-1]*_rdimensions[d-1];
_istride[d] = _istride[d-1]*_simd_layout[d-1];
}
// Use a reduced simd grid
_simd_layout[d] = simd_layout[d];
_rdimensions[d] = _ldimensions[d] / _simd_layout[d]; // this is not checking if this is integer
assert(_rdimensions[d] * _simd_layout[d] == _ldimensions[d]);
assert(_rdimensions[d] > 0);
// all elements of a simd vector must have same checkerboard.
// If Ls vectorised, this must still be the case; e.g. dwf rb5d
if (_simd_layout[d] > 1)
{
if (checker_dim_mask[d])
{
assert((_rdimensions[d] & 0x1) == 0);
}
}
_osites *= _rdimensions[d];
_isites *= _simd_layout[d];
// Addressing support
if (d == 0)
{
_ostride[d] = 1;
_istride[d] = 1;
}
else
{
_ostride[d] = _ostride[d - 1] * _rdimensions[d - 1];
_istride[d] = _istride[d - 1] * _simd_layout[d - 1];
}
}
////////////////////////////////////////////////////////////////////////////////////////////
// subplane information
////////////////////////////////////////////////////////////////////////////////////////////
_slice_block.resize(_ndimension);
_slice_stride.resize(_ndimension);
_slice_nblock.resize(_ndimension);
int block =1;
int nblock=1;
for(int d=0;d<_ndimension;d++) nblock*=_rdimensions[d];
for(int d=0;d<_ndimension;d++){
nblock/=_rdimensions[d];
_slice_block[d] =block;
_slice_stride[d]=_ostride[d]*_rdimensions[d];
_slice_nblock[d]=nblock;
block = block*_rdimensions[d];
int block = 1;
int nblock = 1;
for (int d = 0; d < _ndimension; d++)
nblock *= _rdimensions[d];
for (int d = 0; d < _ndimension; d++)
{
nblock /= _rdimensions[d];
_slice_block[d] = block;
_slice_stride[d] = _ostride[d] * _rdimensions[d];
_slice_nblock[d] = nblock;
block = block * _rdimensions[d];
}
////////////////////////////////////////////////
// Create a checkerboard lookup table
////////////////////////////////////////////////
int rvol = 1;
for(int d=0;d<_ndimension;d++){
rvol=rvol * _rdimensions[d];
for (int d = 0; d < _ndimension; d++)
{
rvol = rvol * _rdimensions[d];
}
_checker_board.resize(rvol);
for(int osite=0;osite<_osites;osite++){
_checker_board[osite] = CheckerBoardFromOindex (osite);
for (int osite = 0; osite < _osites; osite++)
{
_checker_board[osite] = CheckerBoardFromOindex(osite);
}
};
protected:
protected:
virtual int oIndex(std::vector<int> &coor)
{
int idx=0;
for(int d=0;d<_ndimension;d++) {
if( d==_checker_dim ) {
idx+=_ostride[d]*((coor[d]/2)%_rdimensions[d]);
} else {
idx+=_ostride[d]*(coor[d]%_rdimensions[d]);
}
int idx = 0;
for (int d = 0; d < _ndimension; d++)
{
if (d == _checker_dim)
{
idx += _ostride[d] * ((coor[d] / 2) % _rdimensions[d]);
}
else
{
idx += _ostride[d] * (coor[d] % _rdimensions[d]);
}
}
return idx;
};
virtual int iIndex(std::vector<int> &lcoor)
{
int idx=0;
for(int d=0;d<_ndimension;d++) {
if( d==_checker_dim ) {
idx+=_istride[d]*(lcoor[d]/(2*_rdimensions[d]));
} else {
idx+=_istride[d]*(lcoor[d]/_rdimensions[d]);
}
}
return idx;
int idx = 0;
for (int d = 0; d < _ndimension; d++)
{
if (d == _checker_dim)
{
idx += _istride[d] * (lcoor[d] / (2 * _rdimensions[d]));
}
else
{
idx += _istride[d] * (lcoor[d] / _rdimensions[d]);
}
}
return idx;
}
};
}
#endif

61
lib/communicator/Communicator_base.cc
View File

@ -26,6 +26,10 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
*************************************************************************************/
/* END LEGAL */
#include <Grid/GridCore.h>
#include <fcntl.h>
#include <unistd.h>
#include <limits.h>
#include <sys/mman.h>
namespace Grid {
@ -34,7 +38,10 @@ namespace Grid {
///////////////////////////////////////////////////////////////
void * CartesianCommunicator::ShmCommBuf;
uint64_t CartesianCommunicator::MAX_MPI_SHM_BYTES = 128*1024*1024;
CartesianCommunicator::CommunicatorPolicy_t CartesianCommunicator::CommunicatorPolicy= CartesianCommunicator::CommunicatorPolicyConcurrent;
CartesianCommunicator::CommunicatorPolicy_t
CartesianCommunicator::CommunicatorPolicy= CartesianCommunicator::CommunicatorPolicyConcurrent;
int CartesianCommunicator::nCommThreads = -1;
int CartesianCommunicator::Hugepages = 0;
/////////////////////////////////
// Alloc, free shmem region
@ -60,6 +67,7 @@ void CartesianCommunicator::ShmBufferFreeAll(void) {
/////////////////////////////////
// Grid information queries
/////////////////////////////////
int CartesianCommunicator::Dimensions(void) { return _ndimension; };
int CartesianCommunicator::IsBoss(void) { return _processor==0; };
int CartesianCommunicator::BossRank(void) { return 0; };
int CartesianCommunicator::ThisRank(void) { return _processor; };
@ -88,24 +96,43 @@ void CartesianCommunicator::GlobalSumVector(ComplexD *c,int N)
GlobalSumVector((double *)c,2*N);
}
#if !defined( GRID_COMMS_MPI3) && !defined (GRID_COMMS_MPI3L)
#if !defined( GRID_COMMS_MPI3)
int CartesianCommunicator::NodeCount(void) { return ProcessorCount();};
double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
void *xmit,
int xmit_to_rank,
void *recv,
int recv_from_rank,
int bytes)
int CartesianCommunicator::RankCount(void) { return ProcessorCount();};
#endif
#if !defined( GRID_COMMS_MPI3) && !defined (GRID_COMMS_MPIT)
double CartesianCommunicator::StencilSendToRecvFrom( void *xmit,
int xmit_to_rank,
void *recv,
int recv_from_rank,
int bytes, int dir)
{
std::vector<CommsRequest_t> list;
// Discard the "dir"
SendToRecvFromBegin (list,xmit,xmit_to_rank,recv,recv_from_rank,bytes);
SendToRecvFromComplete(list);
return 2.0*bytes;
}
double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
void *xmit,
int xmit_to_rank,
void *recv,
int recv_from_rank,
int bytes, int dir)
{
// Discard the "dir"
SendToRecvFromBegin(list,xmit,xmit_to_rank,recv,recv_from_rank,bytes);
return 2.0*bytes;
}
void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &waitall)
void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &waitall,int dir)
{
SendToRecvFromComplete(waitall);
}
#endif
#if !defined( GRID_COMMS_MPI3)
void CartesianCommunicator::StencilBarrier(void){};
commVector<uint8_t> CartesianCommunicator::ShmBufStorageVector;
@ -119,8 +146,22 @@ void *CartesianCommunicator::ShmBufferTranslate(int rank,void * local_p) {
return NULL;
}
void CartesianCommunicator::ShmInitGeneric(void){
#if 1
int mmap_flag = MAP_SHARED | MAP_ANONYMOUS;
#ifdef MAP_HUGETLB
if ( Hugepages ) mmap_flag |= MAP_HUGETLB;
#endif
ShmCommBuf =(void *) mmap(NULL, MAX_MPI_SHM_BYTES, PROT_READ | PROT_WRITE, mmap_flag, -1, 0);
if (ShmCommBuf == (void *)MAP_FAILED) {
perror("mmap failed ");
exit(EXIT_FAILURE);
}
#else
ShmBufStorageVector.resize(MAX_MPI_SHM_BYTES);
ShmCommBuf=(void *)&ShmBufStorageVector[0];
#endif
bzero(ShmCommBuf,MAX_MPI_SHM_BYTES);
}
#endif

59
lib/communicator/Communicator_base.h
View File

@ -38,7 +38,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
#ifdef GRID_COMMS_MPI3
#include <mpi.h>
#endif
#ifdef GRID_COMMS_MPI3L
#ifdef GRID_COMMS_MPIT
#include <mpi.h>
#endif
#ifdef GRID_COMMS_SHMEM
@ -50,12 +50,24 @@ namespace Grid {
class CartesianCommunicator {
public:
// 65536 ranks per node adequate for now
////////////////////////////////////////////
// Isend/Irecv/Wait, or Sendrecv blocking
////////////////////////////////////////////
enum CommunicatorPolicy_t { CommunicatorPolicyConcurrent, CommunicatorPolicySequential };
static CommunicatorPolicy_t CommunicatorPolicy;
static void SetCommunicatorPolicy(CommunicatorPolicy_t policy ) { CommunicatorPolicy = policy; }
///////////////////////////////////////////
// Up to 65536 ranks per node adequate for now
// 128MB shared memory for comms enought for 48^4 local vol comms
// Give external control (command line override?) of this
static const int MAXLOG2RANKSPERNODE = 16;
static uint64_t MAX_MPI_SHM_BYTES;
///////////////////////////////////////////
static const int MAXLOG2RANKSPERNODE = 16;
static uint64_t MAX_MPI_SHM_BYTES;
static int nCommThreads;
// use explicit huge pages
static int Hugepages;
// Communicator should know nothing of the physics grid, only processor grid.
int _Nprocessors; // How many in all
@ -64,14 +76,18 @@ class CartesianCommunicator {
std::vector<int> _processor_coor; // linear processor coordinate
unsigned long _ndimension;
#if defined (GRID_COMMS_MPI) || defined (GRID_COMMS_MPI3) || defined (GRID_COMMS_MPI3L)
#if defined (GRID_COMMS_MPI) || defined (GRID_COMMS_MPI3) || defined (GRID_COMMS_MPIT)
static MPI_Comm communicator_world;
MPI_Comm communicator;
MPI_Comm communicator;
std::vector<MPI_Comm> communicator_halo;
typedef MPI_Request CommsRequest_t;
#else
typedef int CommsRequest_t;
#endif
////////////////////////////////////////////////////////////////////
// Helper functionality for SHM Windows common to all other impls
////////////////////////////////////////////////////////////////////
@ -117,11 +133,7 @@ class CartesianCommunicator {
/////////////////////////////////
static void * ShmCommBuf;
// Isend/Irecv/Wait, or Sendrecv blocking
enum CommunicatorPolicy_t { CommunicatorPolicyConcurrent, CommunicatorPolicySequential };
static CommunicatorPolicy_t CommunicatorPolicy;
static void SetCommunicatorPolicy(CommunicatorPolicy_t policy ) { CommunicatorPolicy = policy; }
size_t heap_top;
size_t heap_bytes;
@ -148,6 +160,7 @@ class CartesianCommunicator {
int RankFromProcessorCoor(std::vector<int> &coor);
void ProcessorCoorFromRank(int rank,std::vector<int> &coor);
int Dimensions(void) ;
int IsBoss(void) ;
int BossRank(void) ;
int ThisRank(void) ;
@ -155,6 +168,7 @@ class CartesianCommunicator {
const std::vector<int> & ProcessorGrid(void) ;
int ProcessorCount(void) ;
int NodeCount(void) ;
int RankCount(void) ;
////////////////////////////////////////////////////////////////////////////////
// very VERY rarely (Log, serial RNG) we need world without a grid
@ -175,6 +189,8 @@ class CartesianCommunicator {
void GlobalSumVector(ComplexF *c,int N);
void GlobalSum(ComplexD &c);
void GlobalSumVector(ComplexD *c,int N);
void GlobalXOR(uint32_t &);
void GlobalXOR(uint64_t &);
template<class obj> void GlobalSum(obj &o){
typedef typename obj::scalar_type scalar_type;
@ -207,14 +223,21 @@ class CartesianCommunicator {
void SendToRecvFromComplete(std::vector<CommsRequest_t> &waitall);
double StencilSendToRecvFrom(void *xmit,
int xmit_to_rank,
void *recv,
int recv_from_rank,
int bytes,int dir);
double StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
void *xmit,
int xmit_to_rank,
void *recv,
int recv_from_rank,
int bytes);
void *xmit,
int xmit_to_rank,
void *recv,
int recv_from_rank,
int bytes,int dir);
void StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &waitall);
void StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &waitall,int i);
void StencilBarrier(void);
////////////////////////////////////////////////////////////

8
lib/communicator/Communicator_mpi.cc
View File

@ -83,6 +83,14 @@ void CartesianCommunicator::GlobalSum(uint64_t &u){
int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT64_T,MPI_SUM,communicator);
assert(ierr==0);
}
void CartesianCommunicator::GlobalXOR(uint32_t &u){
int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT32_T,MPI_BXOR,communicator);
assert(ierr==0);
}
void CartesianCommunicator::GlobalXOR(uint64_t &u){
int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT64_T,MPI_BXOR,communicator);
assert(ierr==0);
}
void CartesianCommunicator::GlobalSum(float &f){
int ierr=MPI_Allreduce(MPI_IN_PLACE,&f,1,MPI_FLOAT,MPI_SUM,communicator);
assert(ierr==0);

89
lib/communicator/Communicator_mpi3.cc
View File

@ -37,11 +37,12 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
#include <sys/ipc.h>
#include <sys/shm.h>
#include <sys/mman.h>
//#include <zlib.h>
#ifndef SHM_HUGETLB
#define SHM_HUGETLB 04000
#include <zlib.h>
#ifdef HAVE_NUMAIF_H
#include <numaif.h>
#endif
namespace Grid {
///////////////////////////////////////////////////////////////////////////////////////////////////
@ -65,6 +66,7 @@ std::vector<int> CartesianCommunicator::MyGroup;
std::vector<void *> CartesianCommunicator::ShmCommBufs;
int CartesianCommunicator::NodeCount(void) { return GroupSize;};
int CartesianCommunicator::RankCount(void) { return WorldSize;};
#undef FORCE_COMMS
@ -209,10 +211,35 @@ void CartesianCommunicator::Init(int *argc, char ***argv) {
int fd=shm_open(shm_name,O_RDWR|O_CREAT,0666);
if ( fd < 0 ) { perror("failed shm_open"); assert(0); }
ftruncate(fd, size);
int mmap_flag = MAP_SHARED;
#ifdef MAP_HUGETLB
if (Hugepages) mmap_flag |= MAP_HUGETLB;
#endif
void * ptr = mmap(NULL,size, PROT_READ | PROT_WRITE, mmap_flag, fd, 0);
void * ptr = mmap(NULL,size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
if ( ptr == MAP_FAILED ) { perror("failed mmap"); assert(0); }
assert(((uint64_t)ptr&0x3F)==0);
// Experiments; Experiments; Try to force numa domain on the shm segment if we have numaif.h
#if 0
//#ifdef HAVE_NUMAIF_H
int status;
int flags=MPOL_MF_MOVE;
#ifdef KNL
int nodes=1; // numa domain == MCDRAM
// Find out if in SNC2,SNC4 mode ?
#else
int nodes=r; // numa domain == MPI ID
#endif
unsigned long count=1;
for(uint64_t page=0;page<size;page+=4096){
void *pages = (void *) ( page + (uint64_t)ptr );
uint64_t *cow_it = (uint64_t *)pages; *cow_it = 1;
ierr= move_pages(0,count, &pages,&nodes,&status,flags);
if (ierr && (page==0)) perror("numa relocate command failed");
}
#endif
ShmCommBufs[r] =ptr;
}
@ -243,7 +270,11 @@ void CartesianCommunicator::Init(int *argc, char ***argv) {
for(int r=0;r<ShmSize;r++){
size_t size = CartesianCommunicator::MAX_MPI_SHM_BYTES;
key_t key = 0x4545 + r;
if ((shmids[r]= shmget(key,size, SHM_HUGETLB | IPC_CREAT | SHM_R | SHM_W)) < 0) {
int flags = IPC_CREAT | SHM_R | SHM_W;
#ifdef SHM_HUGETLB
flags|=SHM_HUGETLB;
#endif
if ((shmids[r]= shmget(key,size, flags)) < 0) {
int errsv = errno;
printf("Errno %d\n",errsv);
perror("shmget");
@ -374,8 +405,14 @@ CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors)
{
int ierr;
communicator=communicator_world;
_ndimension = processors.size();
communicator_halo.resize (2*_ndimension);
for(int i=0;i<_ndimension*2;i++){
MPI_Comm_dup(communicator,&communicator_halo[i]);
}
////////////////////////////////////////////////////////////////
// Assert power of two shm_size.
////////////////////////////////////////////////////////////////
@ -509,6 +546,14 @@ void CartesianCommunicator::GlobalSum(uint64_t &u){
int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT64_T,MPI_SUM,communicator);
assert(ierr==0);
}
void CartesianCommunicator::GlobalXOR(uint32_t &u){
int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT32_T,MPI_BXOR,communicator);
assert(ierr==0);
}
void CartesianCommunicator::GlobalXOR(uint64_t &u){
int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT64_T,MPI_BXOR,communicator);
assert(ierr==0);
}
void CartesianCommunicator::GlobalSum(float &f){
int ierr=MPI_Allreduce(MPI_IN_PLACE,&f,1,MPI_FLOAT,MPI_SUM,communicator);
assert(ierr==0);
@ -590,13 +635,27 @@ void CartesianCommunicator::SendToRecvFromBegin(std::vector<CommsRequest_t> &lis
}
}
double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
void *xmit,
int dest,
void *recv,
int from,
int bytes)
double CartesianCommunicator::StencilSendToRecvFrom( void *xmit,
int dest,
void *recv,
int from,
int bytes,int dir)
{
std::vector<CommsRequest_t> list;
double offbytes = StencilSendToRecvFromBegin(list,xmit,dest,recv,from,bytes,dir);
StencilSendToRecvFromComplete(list,dir);
return offbytes;
}
double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
void *xmit,
int dest,
void *recv,
int from,
int bytes,int dir)
{
assert(dir < communicator_halo.size());
MPI_Request xrq;
MPI_Request rrq;
@ -615,26 +674,26 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
gfrom = MPI_UNDEFINED;
#endif
if ( gfrom ==MPI_UNDEFINED) {
ierr=MPI_Irecv(recv, bytes, MPI_CHAR,from,from,communicator,&rrq);
ierr=MPI_Irecv(recv, bytes, MPI_CHAR,from,from,communicator_halo[dir],&rrq);
assert(ierr==0);
list.push_back(rrq);
off_node_bytes+=bytes;
}
if ( gdest == MPI_UNDEFINED ) {
ierr =MPI_Isend(xmit, bytes, MPI_CHAR,dest,_processor,communicator,&xrq);
ierr =MPI_Isend(xmit, bytes, MPI_CHAR,dest,_processor,communicator_halo[dir],&xrq);
assert(ierr==0);
list.push_back(xrq);
off_node_bytes+=bytes;
}
if ( CommunicatorPolicy == CommunicatorPolicySequential ) {
this->StencilSendToRecvFromComplete(list);
this->StencilSendToRecvFromComplete(list,dir);
}
return off_node_bytes;
}
void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &waitall)
void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &waitall,int dir)
{
SendToRecvFromComplete(waitall);
}

286
lib/communicator/Communicator_mpit.cc Normal file
View File

@ -0,0 +1,286 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/communicator/Communicator_mpi.cc
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 */
#include <Grid/GridCore.h>
#include <Grid/GridQCDcore.h>
#include <Grid/qcd/action/ActionCore.h>
#include <mpi.h>
namespace Grid {
///////////////////////////////////////////////////////////////////////////////////////////////////
// Info that is setup once and indept of cartesian layout
///////////////////////////////////////////////////////////////////////////////////////////////////
MPI_Comm CartesianCommunicator::communicator_world;
// Should error check all MPI calls.
void CartesianCommunicator::Init(int *argc, char ***argv) {
int flag;
int provided;
MPI_Initialized(&flag); // needed to coexist with other libs apparently
if ( !flag ) {
MPI_Init_thread(argc,argv,MPI_THREAD_MULTIPLE,&provided);
if ( provided != MPI_THREAD_MULTIPLE ) {
QCD::WilsonKernelsStatic::Comms = QCD::WilsonKernelsStatic::CommsThenCompute;
}
}
MPI_Comm_dup (MPI_COMM_WORLD,&communicator_world);
ShmInitGeneric();
}
CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors)
{
_ndimension = processors.size();
std::vector<int> periodic(_ndimension,1);
_Nprocessors=1;
_processors = processors;
_processor_coor.resize(_ndimension);
MPI_Cart_create(communicator_world, _ndimension,&_processors[0],&periodic[0],1,&communicator);
MPI_Comm_rank(communicator,&_processor);
MPI_Cart_coords(communicator,_processor,_ndimension,&_processor_coor[0]);
for(int i=0;i<_ndimension;i++){
_Nprocessors*=_processors[i];
}
communicator_halo.resize (2*_ndimension);
for(int i=0;i<_ndimension*2;i++){
MPI_Comm_dup(communicator,&communicator_halo[i]);
}
int Size;
MPI_Comm_size(communicator,&Size);
assert(Size==_Nprocessors);
}
void CartesianCommunicator::GlobalSum(uint32_t &u){
int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT32_T,MPI_SUM,communicator);
assert(ierr==0);
}
void CartesianCommunicator::GlobalSum(uint64_t &u){
int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT64_T,MPI_SUM,communicator);
assert(ierr==0);
}
void CartesianCommunicator::GlobalXOR(uint32_t &u){
int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT32_T,MPI_BXOR,communicator);
assert(ierr==0);
}
void CartesianCommunicator::GlobalXOR(uint64_t &u){
int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT64_T,MPI_BXOR,communicator);
assert(ierr==0);
}
void CartesianCommunicator::GlobalSum(float &f){
int ierr=MPI_Allreduce(MPI_IN_PLACE,&f,1,MPI_FLOAT,MPI_SUM,communicator);
assert(ierr==0);
}
void CartesianCommunicator::GlobalSumVector(float *f,int N)
{
int ierr=MPI_Allreduce(MPI_IN_PLACE,f,N,MPI_FLOAT,MPI_SUM,communicator);
assert(ierr==0);
}
void CartesianCommunicator::GlobalSum(double &d)
{
int ierr = MPI_Allreduce(MPI_IN_PLACE,&d,1,MPI_DOUBLE,MPI_SUM,communicator);
assert(ierr==0);
}
void CartesianCommunicator::GlobalSumVector(double *d,int N)
{
int ierr = MPI_Allreduce(MPI_IN_PLACE,d,N,MPI_DOUBLE,MPI_SUM,communicator);
assert(ierr==0);
}
void CartesianCommunicator::ShiftedRanks(int dim,int shift,int &source,int &dest)
{
int ierr=MPI_Cart_shift(communicator,dim,shift,&source,&dest);
assert(ierr==0);
}
int CartesianCommunicator::RankFromProcessorCoor(std::vector<int> &coor)
{
int rank;
int ierr=MPI_Cart_rank (communicator, &coor[0], &rank);
assert(ierr==0);
return rank;
}
void CartesianCommunicator::ProcessorCoorFromRank(int rank, std::vector<int> &coor)
{
coor.resize(_ndimension);
int ierr=MPI_Cart_coords (communicator, rank, _ndimension,&coor[0]);
assert(ierr==0);
}
// Basic Halo comms primitive
void CartesianCommunicator::SendToRecvFrom(void *xmit,
int dest,
void *recv,
int from,
int bytes)
{
std::vector<CommsRequest_t> reqs(0);
SendToRecvFromBegin(reqs,xmit,dest,recv,from,bytes);
SendToRecvFromComplete(reqs);
}
void CartesianCommunicator::SendRecvPacket(void *xmit,
void *recv,
int sender,
int receiver,
int bytes)
{
MPI_Status stat;
assert(sender != receiver);
int tag = sender;
if ( _processor == sender ) {
MPI_Send(xmit, bytes, MPI_CHAR,receiver,tag,communicator);
}
if ( _processor == receiver ) {
MPI_Recv(recv, bytes, MPI_CHAR,sender,tag,communicator,&stat);
}
}
// Basic Halo comms primitive
void CartesianCommunicator::SendToRecvFromBegin(std::vector<CommsRequest_t> &list,
void *xmit,
int dest,
void *recv,
int from,
int bytes)
{
int myrank = _processor;
int ierr;
if ( CommunicatorPolicy == CommunicatorPolicyConcurrent ) {
MPI_Request xrq;
MPI_Request rrq;
ierr =MPI_Irecv(recv, bytes, MPI_CHAR,from,from,communicator,&rrq);
ierr|=MPI_Isend(xmit, bytes, MPI_CHAR,dest,_processor,communicator,&xrq);
assert(ierr==0);
list.push_back(xrq);
list.push_back(rrq);
} else {
// Give the CPU to MPI immediately; can use threads to overlap optionally
ierr=MPI_Sendrecv(xmit,bytes,MPI_CHAR,dest,myrank,
recv,bytes,MPI_CHAR,from, from,
communicator,MPI_STATUS_IGNORE);
assert(ierr==0);
}
}
void CartesianCommunicator::SendToRecvFromComplete(std::vector<CommsRequest_t> &list)
{
if ( CommunicatorPolicy == CommunicatorPolicyConcurrent ) {
int nreq=list.size();
std::vector<MPI_Status> status(nreq);
int ierr = MPI_Waitall(nreq,&list[0],&status[0]);
assert(ierr==0);
}
}
void CartesianCommunicator::Barrier(void)
{
int ierr = MPI_Barrier(communicator);
assert(ierr==0);
}
void CartesianCommunicator::Broadcast(int root,void* data, int bytes)
{
int ierr=MPI_Bcast(data,
bytes,
MPI_BYTE,
root,
communicator);
assert(ierr==0);
}
///////////////////////////////////////////////////////
// Should only be used prior to Grid Init finished.
// Check for this?
///////////////////////////////////////////////////////
int CartesianCommunicator::RankWorld(void){
int r;
MPI_Comm_rank(communicator_world,&r);
return r;
}
void CartesianCommunicator::BroadcastWorld(int root,void* data, int bytes)
{
int ierr= MPI_Bcast(data,
bytes,
MPI_BYTE,
root,
communicator_world);
assert(ierr==0);
}
double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
void *xmit,
int xmit_to_rank,
void *recv,
int recv_from_rank,
int bytes,int dir)
{
int myrank = _processor;
int ierr;
assert(dir < communicator_halo.size());
// std::cout << " sending on communicator "<<dir<<" " <<communicator_halo[dir]<<std::endl;
// Give the CPU to MPI immediately; can use threads to overlap optionally
MPI_Request req[2];
MPI_Irecv(recv,bytes,MPI_CHAR,recv_from_rank,recv_from_rank, communicator_halo[dir],&req[1]);
MPI_Isend(xmit,bytes,MPI_CHAR,xmit_to_rank ,myrank , communicator_halo[dir],&req[0]);
list.push_back(req[0]);
list.push_back(req[1]);
return 2.0*bytes;
}
void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &waitall,int dir)
{
int nreq=waitall.size();
MPI_Waitall(nreq, &waitall[0], MPI_STATUSES_IGNORE);
};
double CartesianCommunicator::StencilSendToRecvFrom(void *xmit,
int xmit_to_rank,
void *recv,
int recv_from_rank,
int bytes,int dir)
{
int myrank = _processor;
int ierr;
assert(dir < communicator_halo.size());
// std::cout << " sending on communicator "<<dir<<" " <<communicator_halo[dir]<<std::endl;
// Give the CPU to MPI immediately; can use threads to overlap optionally
MPI_Request req[2];
MPI_Irecv(recv,bytes,MPI_CHAR,recv_from_rank,recv_from_rank, communicator_halo[dir],&req[1]);
MPI_Isend(xmit,bytes,MPI_CHAR,xmit_to_rank ,myrank , communicator_halo[dir],&req[0]);
MPI_Waitall(2, req, MPI_STATUSES_IGNORE);
return 2.0*bytes;
}
}

2
lib/communicator/Communicator_none.cc
View File

@ -59,6 +59,8 @@ void CartesianCommunicator::GlobalSum(double &){}
void CartesianCommunicator::GlobalSum(uint32_t &){}
void CartesianCommunicator::GlobalSum(uint64_t &){}
void CartesianCommunicator::GlobalSumVector(double *,int N){}
void CartesianCommunicator::GlobalXOR(uint32_t &){}
void CartesianCommunicator::GlobalXOR(uint64_t &){}
void CartesianCommunicator::SendRecvPacket(void *xmit,
void *recv,

2
lib/cshift/Cshift.h
View File

@ -42,7 +42,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
#include <Grid/cshift/Cshift_mpi.h>
#endif
#ifdef GRID_COMMS_MPI3L
#ifdef GRID_COMMS_MPIT
#include <Grid/cshift/Cshift_mpi.h>
#endif

245
lib/lattice/Lattice_reduction.h
View File

@ -1,4 +1,4 @@
/*************************************************************************************
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/lattice/Lattice_reduction.h
Copyright (C) 2015
@ -369,71 +369,7 @@ static void sliceMaddVector(Lattice<vobj> &R,std::vector<RealD> &a,const Lattice
}
};
/*
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;
@ -452,7 +388,7 @@ inline GridBase *makeSubSliceGrid(const GridBase *BlockSolverGrid,int Or
}
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)
@ -462,57 +398,170 @@ static void sliceMaddMatrix (Lattice<vobj> &R,Eigen::MatrixXcd &aa,const Lattice
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);
// GridBase *SliceGrid = makeSubSliceGrid(FullGrid,Orthog);
// Lattice<vobj> Xslice(SliceGrid);
// Lattice<vobj> Rslice(SliceGrid);
assert( FullGrid->_simd_layout[Orthog]==1);
int nh = FullGrid->_ndimension;
// int nl = SliceGrid->_ndimension;
int nl = nh-1;
//FIXME package in a convenient iterator
//Should loop over a plane orthogonal to direction "Orthog"
int stride=FullGrid->_slice_stride[Orthog];
int block =FullGrid->_slice_block [Orthog];
int nblock=FullGrid->_slice_nblock[Orthog];
int ostride=FullGrid->_ostride[Orthog];
#pragma omp parallel
{
std::vector<vobj> s_x(Nblock);
#pragma omp for collapse(2)
for(int n=0;n<nblock;n++){
for(int b=0;b<block;b++){
int o = n*stride + b;
for(int i=0;i<Nblock;i++){
s_x[i] = X[o+i*ostride];
}
vobj dot;
for(int i=0;i<Nblock;i++){
dot = Y[o+i*ostride];
for(int j=0;j<Nblock;j++){
dot = dot + s_x[j]*(scale*aa(j,i));
}
R[o+i*ostride]=dot;
}
}}
}
};
template<class vobj>
static void sliceMulMatrix (Lattice<vobj> &R,Eigen::MatrixXcd &aa,const Lattice<vobj> &X,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);
assert( FullGrid->_simd_layout[Orthog]==1);
int nh = FullGrid->_ndimension;
// int nl = SliceGrid->_ndimension;
int nl=1;
//FIXME package in a convenient iterator
//Should loop over a plane orthogonal to direction "Orthog"
int stride=FullGrid->_slice_stride[Orthog];
int block =FullGrid->_slice_block [Orthog];
int nblock=FullGrid->_slice_nblock[Orthog];
int ostride=FullGrid->_ostride[Orthog];
#pragma omp parallel
{
std::vector<vobj> s_x(Nblock);
#pragma omp for collapse(2)
for(int n=0;n<nblock;n++){
for(int b=0;b<block;b++){
int o = n*stride + b;
for(int i=0;i<Nblock;i++){
s_x[i] = X[o+i*ostride];
}
vobj dot;
for(int i=0;i<Nblock;i++){
dot = s_x[0]*(scale*aa(0,i));
for(int j=1;j<Nblock;j++){
dot = dot + s_x[j]*(scale*aa(j,i));
}
R[o+i*ostride]=dot;
}
}}
}
};
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);
// GridBase *SliceGrid = makeSubSliceGrid(FullGrid,Orthog);
int Nblock = FullGrid->GlobalDimensions()[Orthog];
Lattice<vobj> Lslice(SliceGrid);
Lattice<vobj> Rslice(SliceGrid);
// 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);
}
assert( FullGrid->_simd_layout[Orthog]==1);
int nh = FullGrid->_ndimension;
// int nl = SliceGrid->_ndimension;
int nl = nh-1;
//FIXME package in a convenient iterator
//Should loop over a plane orthogonal to direction "Orthog"
int stride=FullGrid->_slice_stride[Orthog];
int block =FullGrid->_slice_block [Orthog];
int nblock=FullGrid->_slice_nblock[Orthog];
int ostride=FullGrid->_ostride[Orthog];
typedef typename vobj::vector_typeD vector_typeD;
#pragma omp parallel
{
std::vector<vobj> Left(Nblock);
std::vector<vobj> Right(Nblock);
Eigen::MatrixXcd mat_thread = Eigen::MatrixXcd::Zero(Nblock,Nblock);
#pragma omp for collapse(2)
for(int n=0;n<nblock;n++){
for(int b=0;b<block;b++){
int o = n*stride + b;
for(int i=0;i<Nblock;i++){
Left [i] = lhs[o+i*ostride];
Right[i] = rhs[o+i*ostride];
}
for(int i=0;i<Nblock;i++){
for(int j=0;j<Nblock;j++){
auto tmp = innerProduct(Left[i],Right[j]);
// vector_typeD rtmp = TensorRemove(tmp);
auto rtmp = TensorRemove(tmp);
mat_thread(i,j) += Reduce(rtmp);
}}
}}
#pragma omp critical
{
mat += mat_thread;
}
}
#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
for(int j=0;j<Nblock;j++){
ComplexD sum = mat(i,j);
FullGrid->GlobalSum(sum);
mat(i,j)=sum;
}}
return;
}

55
lib/lattice/Lattice_transfer.h
View File

@ -551,7 +551,10 @@ void Replicate(Lattice<vobj> &coarse,Lattice<vobj> & fine)
//Copy SIMD-vectorized lattice to array of scalar objects in lexicographic order
template<typename vobj, typename sobj>
typename std::enable_if<isSIMDvectorized<vobj>::value && !isSIMDvectorized<sobj>::value, void>::type unvectorizeToLexOrdArray(std::vector<sobj> &out, const Lattice<vobj> &in){
typename std::enable_if<isSIMDvectorized<vobj>::value && !isSIMDvectorized<sobj>::value, void>::type
unvectorizeToLexOrdArray(std::vector<sobj> &out, const Lattice<vobj> &in)
{
typedef typename vobj::vector_type vtype;
GridBase* in_grid = in._grid;
@ -590,6 +593,54 @@ typename std::enable_if<isSIMDvectorized<vobj>::value && !isSIMDvectorized<sobj>
extract1(in_vobj, out_ptrs, 0);
}
}
//Copy SIMD-vectorized lattice to array of scalar objects in lexicographic order
template<typename vobj, typename sobj>
typename std::enable_if<isSIMDvectorized<vobj>::value
&& !isSIMDvectorized<sobj>::value, void>::type
vectorizeFromLexOrdArray( std::vector<sobj> &in, Lattice<vobj> &out)
{
typedef typename vobj::vector_type vtype;
GridBase* grid = out._grid;
assert(in.size()==grid->lSites());
int ndim = grid->Nd();
int nsimd = vtype::Nsimd();
std::vector<std::vector<int> > icoor(nsimd);
for(int lane=0; lane < nsimd; lane++){
icoor[lane].resize(ndim);
grid->iCoorFromIindex(icoor[lane],lane);
}
parallel_for(uint64_t oidx = 0; oidx < grid->oSites(); oidx++){ //loop over outer index
//Assemble vector of pointers to output elements
std::vector<sobj*> ptrs(nsimd);
std::vector<int> ocoor(ndim);
grid->oCoorFromOindex(ocoor, oidx);
std::vector<int> lcoor(grid->Nd());
for(int lane=0; lane < nsimd; lane++){
for(int mu=0;mu<ndim;mu++){
lcoor[mu] = ocoor[mu] + grid->_rdimensions[mu]*icoor[lane][mu];
}
int lex;
Lexicographic::IndexFromCoor(lcoor, lex, grid->_ldimensions);
ptrs[lane] = &in[lex];
}
//pack from those ptrs
vobj vecobj;
merge1(vecobj, ptrs, 0);
out._odata[oidx] = vecobj;
}
}
//Convert a Lattice from one precision to another
template<class VobjOut, class VobjIn>
@ -615,7 +666,7 @@ void precisionChange(Lattice<VobjOut> &out, const Lattice<VobjIn> &in){
std::vector<SobjOut> in_slex_conv(in_grid->lSites());
unvectorizeToLexOrdArray(in_slex_conv, in);
parallel_for(int out_oidx=0;out_oidx<out_grid->oSites();out_oidx++){
parallel_for(uint64_t out_oidx=0;out_oidx<out_grid->oSites();out_oidx++){
std::vector<int> out_ocoor(ndim);
out_grid->oCoorFromOindex(out_ocoor, out_oidx);

2
lib/log/Log.cc
View File

@ -95,7 +95,7 @@ void GridLogConfigure(std::vector<std::string> &logstreams) {
////////////////////////////////////////////////////////////
void Grid_quiesce_nodes(void) {
int me = 0;
#if defined(GRID_COMMS_MPI) || defined(GRID_COMMS_MPI3) || defined(GRID_COMMS_MPI3L)
#if defined(GRID_COMMS_MPI) || defined(GRID_COMMS_MPI3) || defined(GRID_COMMS_MPIT)
MPI_Comm_rank(MPI_COMM_WORLD, &me);
#endif
#ifdef GRID_COMMS_SHMEM

1192
lib/parallelIO/BinaryIO.h
View File

File diff suppressed because it is too large Load Diff

787
lib/parallelIO/IldgIO.h
View File

@ -27,6 +27,7 @@ directory
#ifndef GRID_ILDG_IO_H
#define GRID_ILDG_IO_H
#ifdef HAVE_LIME
#include <algorithm>
#include <fstream>
#include <iomanip>
@ -37,213 +38,677 @@ directory
#include <sys/utsname.h>
#include <unistd.h>
#ifdef HAVE_LIME
extern "C" { // for linkage
//C-Lime is a must have for this functionality
extern "C" {
#include "lime.h"
}
namespace Grid {
namespace QCD {
inline void ILDGGrid(GridBase *grid, ILDGField &header) {
assert(grid->_ndimension == 4); // emit error if not
header.dimension.resize(4);
header.boundary.resize(4);
for (int d = 0; d < 4; d++) {
header.dimension[d] = grid->_fdimensions[d];
// Read boundary conditions from ... ?
header.boundary[d] = std::string("periodic");
}
}
/////////////////////////////////
// Encode word types as strings
/////////////////////////////////
template<class word> inline std::string ScidacWordMnemonic(void){ return std::string("unknown"); }
template<> inline std::string ScidacWordMnemonic<double> (void){ return std::string("D"); }
template<> inline std::string ScidacWordMnemonic<float> (void){ return std::string("F"); }
template<> inline std::string ScidacWordMnemonic< int32_t>(void){ return std::string("I32_t"); }
template<> inline std::string ScidacWordMnemonic<uint32_t>(void){ return std::string("U32_t"); }
template<> inline std::string ScidacWordMnemonic< int64_t>(void){ return std::string("I64_t"); }
template<> inline std::string ScidacWordMnemonic<uint64_t>(void){ return std::string("U64_t"); }
inline void ILDGChecksum(uint32_t *buf, uint32_t buf_size_bytes,
uint32_t &csum) {
BinaryIO::Uint32Checksum(buf, buf_size_bytes, csum);
}
/////////////////////////////////////////
// Encode a generic tensor as a string
/////////////////////////////////////////
template<class vobj> std::string ScidacRecordTypeString(int &colors, int &spins, int & typesize,int &datacount) {
//////////////////////////////////////////////////////////////////////
// Utilities ; these are QCD aware
//////////////////////////////////////////////////////////////////////
template <class GaugeField>
inline void ILDGStatistics(GaugeField &data, ILDGField &header) {
// How to convert data precision etc...
header.link_trace = Grid::QCD::WilsonLoops<PeriodicGimplR>::linkTrace(data);
header.plaquette = Grid::QCD::WilsonLoops<PeriodicGimplR>::avgPlaquette(data);
// header.polyakov =
}
typedef typename getPrecision<vobj>::real_scalar_type stype;
// Forcing QCD here
template <class fobj, class sobj>
struct ILDGMunger {
void operator()(fobj &in, sobj &out, uint32_t &csum) {
for (int mu = 0; mu < 4; mu++) {
for (int i = 0; i < 3; i++) {
for (int j = 0; j < 3; j++) {
out(mu)()(i, j) = in(mu)()(i, j);
}
}
}
ILDGChecksum((uint32_t *)&in, sizeof(in), csum);
};
};
int _ColourN = indexRank<ColourIndex,vobj>();
int _ColourScalar = isScalar<ColourIndex,vobj>();
int _ColourVector = isVector<ColourIndex,vobj>();
int _ColourMatrix = isMatrix<ColourIndex,vobj>();
template <class fobj, class sobj>
struct ILDGUnmunger {
void operator()(sobj &in, fobj &out, uint32_t &csum) {
for (int mu = 0; mu < 4; mu++) {
for (int i = 0; i < 3; i++) {
for (int j = 0; j < 3; j++) {
out(mu)()(i, j) = in(mu)()(i, j);
}
}
}
ILDGChecksum((uint32_t *)&out, sizeof(out), csum);
};
};
int _SpinN = indexRank<SpinIndex,vobj>();
int _SpinScalar = isScalar<SpinIndex,vobj>();
int _SpinVector = isVector<SpinIndex,vobj>();
int _SpinMatrix = isMatrix<SpinIndex,vobj>();
////////////////////////////////////////////////////////////////////////////////
// Write and read from fstream; compute header offset for payload
////////////////////////////////////////////////////////////////////////////////
enum ILDGstate {ILDGread, ILDGwrite};
int _LorentzN = indexRank<LorentzIndex,vobj>();
int _LorentzScalar = isScalar<LorentzIndex,vobj>();
int _LorentzVector = isVector<LorentzIndex,vobj>();
int _LorentzMatrix = isMatrix<LorentzIndex,vobj>();
class ILDGIO : public BinaryIO {
FILE *File;
LimeWriter *LimeW;
LimeRecordHeader *LimeHeader;
LimeReader *LimeR;
std::string filename;
std::stringstream stream;
stream << "GRID_";
stream << ScidacWordMnemonic<stype>();
// std::cout << " Lorentz N/S/V/M : " << _LorentzN<<" "<<_LorentzScalar<<"/"<<_LorentzVector<<"/"<<_LorentzMatrix<<std::endl;
// std::cout << " Spin N/S/V/M : " << _SpinN <<" "<<_SpinScalar <<"/"<<_SpinVector <<"/"<<_SpinMatrix<<std::endl;
// std::cout << " Colour N/S/V/M : " << _ColourN <<" "<<_ColourScalar <<"/"<<_ColourVector <<"/"<<_ColourMatrix<<std::endl;
if ( _LorentzVector ) stream << "_LorentzVector"<<_LorentzN;
if ( _LorentzMatrix ) stream << "_LorentzMatrix"<<_LorentzN;
if ( _SpinVector ) stream << "_SpinVector"<<_SpinN;
if ( _SpinMatrix ) stream << "_SpinMatrix"<<_SpinN;
if ( _ColourVector ) stream << "_ColourVector"<<_ColourN;
if ( _ColourMatrix ) stream << "_ColourMatrix"<<_ColourN;
if ( _ColourScalar && _LorentzScalar && _SpinScalar ) stream << "_Complex";
typesize = sizeof(typename vobj::scalar_type);
if ( _ColourMatrix ) typesize*= _ColourN*_ColourN;
else typesize*= _ColourN;
if ( _SpinMatrix ) typesize*= _SpinN*_SpinN;
else typesize*= _SpinN;
colors = _ColourN;
spins = _SpinN;
datacount = _LorentzN;
return stream.str();
}
template<class vobj> std::string ScidacRecordTypeString(Lattice<vobj> & lat,int &colors, int &spins, int & typesize,int &datacount) {
return ScidacRecordTypeString<vobj>(colors,spins,typesize,datacount);
};
////////////////////////////////////////////////////////////
// Helper to fill out metadata
////////////////////////////////////////////////////////////
template<class vobj> void ScidacMetaData(Lattice<vobj> & field,
FieldMetaData &header,
scidacRecord & _scidacRecord,
scidacFile & _scidacFile)
{
typedef typename getPrecision<vobj>::real_scalar_type stype;
/////////////////////////////////////
// Pull Grid's metadata
/////////////////////////////////////
PrepareMetaData(field,header);
/////////////////////////////////////
// Scidac Private File structure
/////////////////////////////////////
_scidacFile = scidacFile(field._grid);
/////////////////////////////////////
// Scidac Private Record structure
/////////////////////////////////////
scidacRecord sr;
sr.datatype = ScidacRecordTypeString(field,sr.colors,sr.spins,sr.typesize,sr.datacount);
sr.date = header.creation_date;
sr.precision = ScidacWordMnemonic<stype>();
sr.recordtype = GRID_IO_FIELD;
_scidacRecord = sr;
std::cout << GridLogMessage << "Build SciDAC datatype " <<sr.datatype<<std::endl;
}
///////////////////////////////////////////////////////
// Scidac checksum
///////////////////////////////////////////////////////
static int scidacChecksumVerify(scidacChecksum &scidacChecksum_,uint32_t scidac_csuma,uint32_t scidac_csumb)
{
uint32_t scidac_checksuma = stoull(scidacChecksum_.suma,0,16);
uint32_t scidac_checksumb = stoull(scidacChecksum_.sumb,0,16);
if ( scidac_csuma !=scidac_checksuma) return 0;
if ( scidac_csumb !=scidac_checksumb) return 0;
return 1;
}
////////////////////////////////////////////////////////////////////////////////////
// Lime, ILDG and Scidac I/O classes
////////////////////////////////////////////////////////////////////////////////////
class GridLimeReader : public BinaryIO {
public:
ILDGIO(std::string file, ILDGstate RW) {
filename = file;
if (RW == ILDGwrite){
File = fopen(file.c_str(), "w");
// check if opened correctly
///////////////////////////////////////////////////
// FIXME: format for RNG? Now just binary out instead
///////////////////////////////////////////////////
LimeW = limeCreateWriter(File);
} else {
File = fopen(file.c_str(), "r");
// check if opened correctly
FILE *File;
LimeReader *LimeR;
std::string filename;
LimeR = limeCreateReader(File);
/////////////////////////////////////////////
// Open the file
/////////////////////////////////////////////
void open(std::string &_filename)
{
filename= _filename;
File = fopen(filename.c_str(), "r");
LimeR = limeCreateReader(File);
}
/////////////////////////////////////////////
// Close the file
/////////////////////////////////////////////
void close(void){
fclose(File);
// limeDestroyReader(LimeR);
}
////////////////////////////////////////////
// Read a generic lattice field and verify checksum
////////////////////////////////////////////
template<class vobj>
void readLimeLatticeBinaryObject(Lattice<vobj> &field,std::string record_name)
{
typedef typename vobj::scalar_object sobj;
scidacChecksum scidacChecksum_;
uint32_t nersc_csum,scidac_csuma,scidac_csumb;
std::string format = getFormatString<vobj>();
while ( limeReaderNextRecord(LimeR) == LIME_SUCCESS ) {
std::cout << GridLogMessage << limeReaderType(LimeR) <<std::endl;
if ( strncmp(limeReaderType(LimeR), record_name.c_str(),strlen(record_name.c_str()) ) ) {
off_t offset= ftell(File);
BinarySimpleMunger<sobj,sobj> munge;
BinaryIO::readLatticeObject< sobj, sobj >(field, filename, munge, offset, format,nersc_csum,scidac_csuma,scidac_csumb);
/////////////////////////////////////////////
// Insist checksum is next record
/////////////////////////////////////////////
readLimeObject(scidacChecksum_,std::string("scidacChecksum"),record_name);
/////////////////////////////////////////////
// Verify checksums
/////////////////////////////////////////////
scidacChecksumVerify(scidacChecksum_,scidac_csuma,scidac_csumb);
return;
}
}
}
////////////////////////////////////////////
// Read a generic serialisable object
////////////////////////////////////////////
template<class serialisable_object>
void readLimeObject(serialisable_object &object,std::string object_name,std::string record_name)
{
std::string xmlstring;
// should this be a do while; can we miss a first record??
while ( limeReaderNextRecord(LimeR) == LIME_SUCCESS ) {
~ILDGIO() { fclose(File); }
uint64_t nbytes = limeReaderBytes(LimeR);//size of this record (configuration)
int createHeader(std::string message, int MB, int ME, size_t PayloadSize, LimeWriter* L){
if ( strncmp(limeReaderType(LimeR), record_name.c_str(),strlen(record_name.c_str()) ) ) {
std::vector<char> xmlc(nbytes+1,'\0');
limeReaderReadData((void *)&xmlc[0], &nbytes, LimeR);
XmlReader RD(&xmlc[0],"");
read(RD,object_name,object);
return;
}
}
assert(0);
}
};
class GridLimeWriter : public BinaryIO {
public:
///////////////////////////////////////////////////
// FIXME: format for RNG? Now just binary out instead
///////////////////////////////////////////////////
FILE *File;
LimeWriter *LimeW;
std::string filename;
void open(std::string &_filename) {
filename= _filename;
File = fopen(filename.c_str(), "w");
LimeW = limeCreateWriter(File); assert(LimeW != NULL );
}
/////////////////////////////////////////////
// Close the file
/////////////////////////////////////////////
void close(void) {
fclose(File);
// limeDestroyWriter(LimeW);
}
///////////////////////////////////////////////////////
// Lime utility functions
///////////////////////////////////////////////////////
int createLimeRecordHeader(std::string message, int MB, int ME, size_t PayloadSize)
{
LimeRecordHeader *h;
h = limeCreateHeader(MB, ME, const_cast<char *>(message.c_str()), PayloadSize);
int status = limeWriteRecordHeader(h, L);
if (status < 0) {
std::cerr << "ILDG Header error\n";
return status;
}
assert(limeWriteRecordHeader(h, LimeW) >= 0);
limeDestroyHeader(h);
return LIME_SUCCESS;
}
////////////////////////////////////////////
// Write a generic serialisable object
////////////////////////////////////////////
template<class serialisable_object>
void writeLimeObject(int MB,int ME,serialisable_object &object,std::string object_name,std::string record_name)
{
std::string xmlstring;
{
XmlWriter WR("","");
write(WR,object_name,object);
xmlstring = WR.XmlString();
}
uint64_t nbytes = xmlstring.size();
int err;
LimeRecordHeader *h = limeCreateHeader(MB, ME,(char *)record_name.c_str(), nbytes); assert(h!= NULL);
unsigned int writeHeader(ILDGField &header) {
// write header in LIME
n_uint64_t nbytes;
int MB_flag = 1, ME_flag = 0;
err=limeWriteRecordHeader(h, LimeW); assert(err>=0);
err=limeWriteRecordData(&xmlstring[0], &nbytes, LimeW); assert(err>=0);
err=limeWriterCloseRecord(LimeW); assert(err>=0);
limeDestroyHeader(h);
}
////////////////////////////////////////////
// Write a generic lattice field and csum
////////////////////////////////////////////
template<class vobj>
void writeLimeLatticeBinaryObject(Lattice<vobj> &field,std::string record_name)
{
////////////////////////////////////////////
// Create record header
////////////////////////////////////////////
typedef typename vobj::scalar_object sobj;
int err;
uint32_t nersc_csum,scidac_csuma,scidac_csumb;
uint64_t PayloadSize = sizeof(sobj) * field._grid->_gsites;
createLimeRecordHeader(record_name, 0, 0, PayloadSize);
char message[] = "ildg-format";
nbytes = strlen(message);
LimeHeader = limeCreateHeader(MB_flag, ME_flag, message, nbytes);
limeWriteRecordHeader(LimeHeader, LimeW);
limeDestroyHeader(LimeHeader);
// save the xml header here
// use the xml_writer to c++ streams in pugixml
// and convert to char message
limeWriteRecordData(message, &nbytes, LimeW);
limeWriterCloseRecord(LimeW);
////////////////////////////////////////////////////////////////////
// NB: FILE and iostream are jointly writing disjoint sequences in the
// the same file through different file handles (integer units).
//
// These are both buffered, so why I think this code is right is as follows.
//
// i) write record header to FILE *File, telegraphing the size.
// ii) ftell reads the offset from FILE *File .
// iii) iostream / MPI Open independently seek this offset. Write sequence direct to disk.
// Closes iostream and flushes.
// iv) fseek on FILE * to end of this disjoint section.
// v) Continue writing scidac record.
////////////////////////////////////////////////////////////////////
off_t offset = ftell(File);
std::string format = getFormatString<vobj>();
BinarySimpleMunger<sobj,sobj> munge;
BinaryIO::writeLatticeObject<vobj,sobj>(field, filename, munge, offset, format,nersc_csum,scidac_csuma,scidac_csumb);
err=limeWriterCloseRecord(LimeW); assert(err>=0);
////////////////////////////////////////
// Write checksum element, propagaing forward from the BinaryIO
// Always pair a checksum with a binary object, and close message
////////////////////////////////////////
scidacChecksum checksum;
std::stringstream streama; streama << std::hex << scidac_csuma;
std::stringstream streamb; streamb << std::hex << scidac_csumb;
checksum.suma= streama.str();
checksum.sumb= streamb.str();
std::cout << GridLogMessage<<" writing scidac checksums "<<std::hex<<scidac_csuma<<"/"<<scidac_csumb<<std::dec<<std::endl;
writeLimeObject(0,1,checksum,std::string("scidacChecksum" ),std::string(SCIDAC_CHECKSUM));
}
};
return 0;
}
unsigned int readHeader(ILDGField &header) {
return 0;
class ScidacWriter : public GridLimeWriter {
public:
template<class SerialisableUserFile>
void writeScidacFileRecord(GridBase *grid,SerialisableUserFile &_userFile)
{
scidacFile _scidacFile(grid);
writeLimeObject(1,0,_scidacFile,_scidacFile.SerialisableClassName(),std::string(SCIDAC_PRIVATE_FILE_XML));
writeLimeObject(0,1,_userFile,_userFile.SerialisableClassName(),std::string(SCIDAC_FILE_XML));
}
////////////////////////////////////////////////
// Write generic lattice field in scidac format
////////////////////////////////////////////////
template <class vobj, class userRecord>
void writeScidacFieldRecord(Lattice<vobj> &field,userRecord _userRecord)
{
typedef typename vobj::scalar_object sobj;
uint64_t nbytes;
GridBase * grid = field._grid;
////////////////////////////////////////
// fill the Grid header
////////////////////////////////////////
FieldMetaData header;
scidacRecord _scidacRecord;
scidacFile _scidacFile;
ScidacMetaData(field,header,_scidacRecord,_scidacFile);
//////////////////////////////////////////////
// Fill the Lime file record by record
//////////////////////////////////////////////
writeLimeObject(1,0,header ,std::string("FieldMetaData"),std::string(GRID_FORMAT)); // Open message
writeLimeObject(0,0,_userRecord,_userRecord.SerialisableClassName(),std::string(SCIDAC_RECORD_XML));
writeLimeObject(0,0,_scidacRecord,_scidacRecord.SerialisableClassName(),std::string(SCIDAC_PRIVATE_RECORD_XML));
writeLimeLatticeBinaryObject(field,std::string(ILDG_BINARY_DATA)); // Closes message with checksum
}
};
class IldgWriter : public ScidacWriter {
public:
///////////////////////////////////
// A little helper
///////////////////////////////////
void writeLimeIldgLFN(std::string &LFN)
{
uint64_t PayloadSize = LFN.size();
int err;
createLimeRecordHeader(ILDG_DATA_LFN, 0 , 0, PayloadSize);
err=limeWriteRecordData(const_cast<char*>(LFN.c_str()), &PayloadSize,LimeW); assert(err>=0);
err=limeWriterCloseRecord(LimeW); assert(err>=0);
}
////////////////////////////////////////////////////////////////
// Special ILDG operations ; gauge configs only.
// Don't require scidac records EXCEPT checksum
// Use Grid MetaData object if present.
////////////////////////////////////////////////////////////////
template <class vsimd>
uint32_t readConfiguration(Lattice<iLorentzColourMatrix<vsimd> > &Umu) {
typedef Lattice<iLorentzColourMatrix<vsimd> > GaugeField;
typedef LorentzColourMatrixD sobjd;
typedef LorentzColourMatrixF sobjf;
typedef iLorentzColourMatrix<vsimd> itype;
typedef LorentzColourMatrix sobj;
GridBase *grid = Umu._grid;
ILDGField header;
readHeader(header);
// now just the conf, ignore the header
std::string format = std::string("IEEE64BIG");
do {limeReaderNextRecord(LimeR);}
while (strncmp(limeReaderType(LimeR), "ildg-binary-data",16));
n_uint64_t nbytes = limeReaderBytes(LimeR);//size of this record (configuration)
ILDGtype ILDGt(true, LimeR);
// this is special for double prec data, just for the moment
uint32_t csum = BinaryIO::readObjectParallel< itype, sobjd >(
Umu, filename, ILDGMunger<sobjd, sobj>(), 0, format, ILDGt);
// Check configuration
// todo
return csum;
}
template <class vsimd>
uint32_t writeConfiguration(Lattice<iLorentzColourMatrix<vsimd> > &Umu, std::string format) {
void writeConfiguration(Lattice<iLorentzColourMatrix<vsimd> > &Umu,int sequence,std::string LFN,std::string description)
{
GridBase * grid = Umu._grid;
typedef Lattice<iLorentzColourMatrix<vsimd> > GaugeField;
typedef iLorentzColourMatrix<vsimd> vobj;
typedef typename vobj::scalar_object sobj;
typedef LorentzColourMatrixD fobj;
ILDGField header;
// fill the header
header.floating_point = format;
uint64_t nbytes;
ILDGUnmunger<fobj, sobj> munge;
unsigned int offset = writeHeader(header);
////////////////////////////////////////
// fill the Grid header
////////////////////////////////////////
FieldMetaData header;
scidacRecord _scidacRecord;
scidacFile _scidacFile;
BinaryIO::Uint32Checksum<vobj, fobj>(Umu, munge, header.checksum);
ScidacMetaData(Umu,header,_scidacRecord,_scidacFile);
// Write data record header
n_uint64_t PayloadSize = sizeof(fobj) * Umu._grid->_gsites;
createHeader("ildg-binary-data", 0, 1, PayloadSize, LimeW);
std::string format = header.floating_point;
header.ensemble_id = description;
header.ensemble_label = description;
header.sequence_number = sequence;
header.ildg_lfn = LFN;
ILDGtype ILDGt(true, LimeW);
uint32_t csum = BinaryIO::writeObjectParallel<vobj, fobj>(
Umu, filename, munge, 0, header.floating_point, ILDGt);
assert ( (format == std::string("IEEE32BIG"))
||(format == std::string("IEEE64BIG")) );
limeWriterCloseRecord(LimeW);
//////////////////////////////////////////////////////
// Fill ILDG header data struct
//////////////////////////////////////////////////////
ildgFormat ildgfmt ;
ildgfmt.field = std::string("su3gauge");
// Last record
// the logical file name LNF
// look into documentation on how to generate this string
std::string LNF = "empty";
if ( format == std::string("IEEE32BIG") ) {
ildgfmt.precision = 32;
} else {
ildgfmt.precision = 64;
}
ildgfmt.version = 1.0;
ildgfmt.lx = header.dimension[0];
ildgfmt.ly = header.dimension[1];
ildgfmt.lz = header.dimension[2];
ildgfmt.lt = header.dimension[3];
assert(header.nd==4);
assert(header.nd==header.dimension.size());
//////////////////////////////////////////////////////////////////////////////
// Fill the USQCD info field
//////////////////////////////////////////////////////////////////////////////
usqcdInfo info;
info.version=1.0;
info.plaq = header.plaquette;
info.linktr = header.link_trace;
PayloadSize = sizeof(LNF);
createHeader("ildg-binary-lfn", 1 , 1, PayloadSize, LimeW);
limeWriteRecordData(const_cast<char*>(LNF.c_str()), &PayloadSize, LimeW);
limeWriterCloseRecord(LimeW);
return csum;
std::cout << GridLogMessage << " Writing config; IldgIO "<<std::endl;
//////////////////////////////////////////////
// Fill the Lime file record by record
//////////////////////////////////////////////
writeLimeObject(1,0,header ,std::string("FieldMetaData"),std::string(GRID_FORMAT)); // Open message
writeLimeObject(0,0,_scidacFile,_scidacFile.SerialisableClassName(),std::string(SCIDAC_PRIVATE_FILE_XML));
writeLimeObject(0,1,info,info.SerialisableClassName(),std::string(SCIDAC_FILE_XML));
writeLimeObject(1,0,_scidacRecord,_scidacRecord.SerialisableClassName(),std::string(SCIDAC_PRIVATE_RECORD_XML));
writeLimeObject(0,0,info,info.SerialisableClassName(),std::string(SCIDAC_RECORD_XML));
writeLimeObject(0,0,ildgfmt,std::string("ildgFormat") ,std::string(ILDG_FORMAT)); // rec
writeLimeIldgLFN(header.ildg_lfn); // rec
writeLimeLatticeBinaryObject(Umu,std::string(ILDG_BINARY_DATA)); // Closes message with checksum
// limeDestroyWriter(LimeW);
fclose(File);
}
// format for RNG? Now just binary out
};
}
}
class IldgReader : public GridLimeReader {
public:
////////////////////////////////////////////////////////////////
// Read either Grid/SciDAC/ILDG configuration
// Don't require scidac records EXCEPT checksum
// Use Grid MetaData object if present.
// Else use ILDG MetaData object if present.
// Else use SciDAC MetaData object if present.
////////////////////////////////////////////////////////////////
template <class vsimd>
void readConfiguration(Lattice<iLorentzColourMatrix<vsimd> > &Umu, FieldMetaData &FieldMetaData_) {
typedef Lattice<iLorentzColourMatrix<vsimd> > GaugeField;
typedef typename GaugeField::vector_object vobj;
typedef typename vobj::scalar_object sobj;
typedef LorentzColourMatrixF fobj;
typedef LorentzColourMatrixD dobj;
GridBase *grid = Umu._grid;
std::vector<int> dims = Umu._grid->FullDimensions();
assert(dims.size()==4);
// Metadata holders
ildgFormat ildgFormat_ ;
std::string ildgLFN_ ;
scidacChecksum scidacChecksum_;
usqcdInfo usqcdInfo_ ;
// track what we read from file
int found_ildgFormat =0;
int found_ildgLFN =0;
int found_scidacChecksum=0;
int found_usqcdInfo =0;
int found_ildgBinary =0;
int found_FieldMetaData =0;
uint32_t nersc_csum;
uint32_t scidac_csuma;
uint32_t scidac_csumb;
// Binary format
std::string format;
//////////////////////////////////////////////////////////////////////////
// Loop over all records
// -- Order is poorly guaranteed except ILDG header preceeds binary section.
// -- Run like an event loop.
// -- Impose trust hierarchy. Grid takes precedence & look for ILDG, and failing
// that Scidac.
// -- Insist on Scidac checksum record.
//////////////////////////////////////////////////////////////////////////
while ( limeReaderNextRecord(LimeR) == LIME_SUCCESS ) {
uint64_t nbytes = limeReaderBytes(LimeR);//size of this record (configuration)
//////////////////////////////////////////////////////////////////
// If not BINARY_DATA read a string and parse
//////////////////////////////////////////////////////////////////
if ( strncmp(limeReaderType(LimeR), ILDG_BINARY_DATA,strlen(ILDG_BINARY_DATA) ) ) {
// Copy out the string
std::vector<char> xmlc(nbytes+1,'\0');
limeReaderReadData((void *)&xmlc[0], &nbytes, LimeR);
std::cout << GridLogMessage<< "Non binary record :" <<limeReaderType(LimeR) <<std::endl; //<<"\n"<<(&xmlc[0])<<std::endl;
//////////////////////////////////
// ILDG format record
if ( !strncmp(limeReaderType(LimeR), ILDG_FORMAT,strlen(ILDG_FORMAT)) ) {
XmlReader RD(&xmlc[0],"");
read(RD,"ildgFormat",ildgFormat_);
if ( ildgFormat_.precision == 64 ) format = std::string("IEEE64BIG");
if ( ildgFormat_.precision == 32 ) format = std::string("IEEE32BIG");
assert( ildgFormat_.lx == dims[0]);
assert( ildgFormat_.ly == dims[1]);
assert( ildgFormat_.lz == dims[2]);
assert( ildgFormat_.lt == dims[3]);
found_ildgFormat = 1;
}
if ( !strncmp(limeReaderType(LimeR), ILDG_DATA_LFN,strlen(ILDG_DATA_LFN)) ) {
FieldMetaData_.ildg_lfn = std::string(&xmlc[0]);
found_ildgLFN = 1;
}
if ( !strncmp(limeReaderType(LimeR), GRID_FORMAT,strlen(ILDG_FORMAT)) ) {
XmlReader RD(&xmlc[0],"");
read(RD,"FieldMetaData",FieldMetaData_);
format = FieldMetaData_.floating_point;
assert(FieldMetaData_.dimension[0] == dims[0]);
assert(FieldMetaData_.dimension[1] == dims[1]);
assert(FieldMetaData_.dimension[2] == dims[2]);
assert(FieldMetaData_.dimension[3] == dims[3]);
found_FieldMetaData = 1;
}
if ( !strncmp(limeReaderType(LimeR), SCIDAC_RECORD_XML,strlen(SCIDAC_RECORD_XML)) ) {
std::string xmls(&xmlc[0]);
// is it a USQCD info field
if ( xmls.find(std::string("usqcdInfo")) != std::string::npos ) {
std::cout << GridLogMessage<<"...found a usqcdInfo field"<<std::endl;
XmlReader RD(&xmlc[0],"");
read(RD,"usqcdInfo",usqcdInfo_);
found_usqcdInfo = 1;
}
}
if ( !strncmp(limeReaderType(LimeR), SCIDAC_CHECKSUM,strlen(SCIDAC_CHECKSUM)) ) {
XmlReader RD(&xmlc[0],"");
read(RD,"scidacChecksum",scidacChecksum_);
found_scidacChecksum = 1;
}
} else {
/////////////////////////////////
// Binary data
/////////////////////////////////
std::cout << GridLogMessage << "ILDG Binary record found : " ILDG_BINARY_DATA << std::endl;
off_t offset= ftell(File);
if ( format == std::string("IEEE64BIG") ) {
GaugeSimpleMunger<dobj, sobj> munge;
BinaryIO::readLatticeObject< vobj, dobj >(Umu, filename, munge, offset, format,nersc_csum,scidac_csuma,scidac_csumb);
} else {
GaugeSimpleMunger<fobj, sobj> munge;
BinaryIO::readLatticeObject< vobj, fobj >(Umu, filename, munge, offset, format,nersc_csum,scidac_csuma,scidac_csumb);
}
found_ildgBinary = 1;
}
}
//////////////////////////////////////////////////////
// Minimally must find binary segment and checksum
// Since this is an ILDG reader require ILDG format
//////////////////////////////////////////////////////
assert(found_ildgBinary);
assert(found_ildgFormat);
assert(found_scidacChecksum);
// Must find something with the lattice dimensions
assert(found_FieldMetaData||found_ildgFormat);
if ( found_FieldMetaData ) {
std::cout << GridLogMessage<<"Grid MetaData was record found: configuration was probably written by Grid ! Yay ! "<<std::endl;
} else {
assert(found_ildgFormat);
assert ( ildgFormat_.field == std::string("su3gauge") );
///////////////////////////////////////////////////////////////////////////////////////
// Populate our Grid metadata as best we can
///////////////////////////////////////////////////////////////////////////////////////
std::ostringstream vers; vers << ildgFormat_.version;
FieldMetaData_.hdr_version = vers.str();
FieldMetaData_.data_type = std::string("4D_SU3_GAUGE_3X3");
FieldMetaData_.nd=4;
FieldMetaData_.dimension.resize(4);
FieldMetaData_.dimension[0] = ildgFormat_.lx ;
FieldMetaData_.dimension[1] = ildgFormat_.ly ;
FieldMetaData_.dimension[2] = ildgFormat_.lz ;
FieldMetaData_.dimension[3] = ildgFormat_.lt ;
if ( found_usqcdInfo ) {
FieldMetaData_.plaquette = usqcdInfo_.plaq;
FieldMetaData_.link_trace= usqcdInfo_.linktr;
std::cout << GridLogMessage <<"This configuration was probably written by USQCD "<<std::endl;
std::cout << GridLogMessage <<"USQCD xml record Plaquette : "<<FieldMetaData_.plaquette<<std::endl;
std::cout << GridLogMessage <<"USQCD xml record LinkTrace : "<<FieldMetaData_.link_trace<<std::endl;
} else {
FieldMetaData_.plaquette = 0.0;
FieldMetaData_.link_trace= 0.0;
std::cout << GridLogWarning << "This configuration is unsafe with no plaquette records that can verify it !!! "<<std::endl;
}
}
////////////////////////////////////////////////////////////
// Really really want to mandate a scidac checksum
////////////////////////////////////////////////////////////
if ( found_scidacChecksum ) {
FieldMetaData_.scidac_checksuma = stoull(scidacChecksum_.suma,0,16);
FieldMetaData_.scidac_checksumb = stoull(scidacChecksum_.sumb,0,16);
scidacChecksumVerify(scidacChecksum_,scidac_csuma,scidac_csumb);
assert( scidac_csuma ==FieldMetaData_.scidac_checksuma);
assert( scidac_csumb ==FieldMetaData_.scidac_checksumb);
std::cout << GridLogMessage<<"SciDAC checksums match " << std::endl;
} else {
std::cout << GridLogWarning<<"SciDAC checksums not found. This is unsafe. " << std::endl;
assert(0); // Can I insist always checksum ?
}
if ( found_FieldMetaData || found_usqcdInfo ) {
FieldMetaData checker;
GaugeStatistics(Umu,checker);
assert(fabs(checker.plaquette - FieldMetaData_.plaquette )<1.0e-5);
assert(fabs(checker.link_trace - FieldMetaData_.link_trace)<1.0e-5);
std::cout << GridLogMessage<<"Plaquette and link trace match " << std::endl;
}
}
};
}}
//HAVE_LIME
#endif

225
lib/parallelIO/IldgIOtypes.h
View File

@ -34,47 +34,198 @@ extern "C" { // for linkage
namespace Grid {
struct ILDGtype {
bool is_ILDG;
LimeWriter* LW;
LimeReader* LR;
/////////////////////////////////////////////////////////////////////////////////
// Data representation of records that enter ILDG and SciDac formats
/////////////////////////////////////////////////////////////////////////////////
ILDGtype(bool is, LimeWriter* L) : is_ILDG(is), LW(L), LR(NULL) {}
ILDGtype(bool is, LimeReader* L) : is_ILDG(is), LW(NULL), LR(L) {}
ILDGtype() : is_ILDG(false), LW(NULL), LR(NULL) {}
};
#define GRID_FORMAT "grid-format"
#define ILDG_FORMAT "ildg-format"
#define ILDG_BINARY_DATA "ildg-binary-data"
#define ILDG_DATA_LFN "ildg-data-lfn"
#define SCIDAC_CHECKSUM "scidac-checksum"
#define SCIDAC_PRIVATE_FILE_XML "scidac-private-file-xml"
#define SCIDAC_FILE_XML "scidac-file-xml"
#define SCIDAC_PRIVATE_RECORD_XML "scidac-private-record-xml"
#define SCIDAC_RECORD_XML "scidac-record-xml"
#define SCIDAC_BINARY_DATA "scidac-binary-data"
// Unused SCIDAC records names; could move to support this functionality
#define SCIDAC_SITELIST "scidac-sitelist"
class ILDGField {
////////////////////////////////////////////////////////////
const int GRID_IO_SINGLEFILE = 0; // hardcode lift from QIO compat
const int GRID_IO_MULTIFILE = 1; // hardcode lift from QIO compat
const int GRID_IO_FIELD = 0; // hardcode lift from QIO compat
const int GRID_IO_GLOBAL = 1; // hardcode lift from QIO compat
////////////////////////////////////////////////////////////
/////////////////////////////////////////////////////////////////////////////////
// QIO uses mandatory "private" records fixed format
// Private is in principle "opaque" however it can't be changed now because that would break existing
// file compatability, so should be correct to assume the undocumented but defacto file structure.
/////////////////////////////////////////////////////////////////////////////////
////////////////////////
// Scidac private file xml
// <?xml version="1.0" encoding="UTF-8"?><scidacFile><version>1.1</version><spacetime>4</spacetime><dims>16 16 16 32 </dims><volfmt>0</volfmt></scidacFile>
////////////////////////
struct scidacFile : Serializable {
public:
// header strings (not in order)
std::vector<int> dimension;
std::vector<std::string> boundary;
int data_start;
std::string hdr_version;
std::string storage_format;
// Checks on data
double link_trace;
double plaquette;
uint32_t checksum;
unsigned int sequence_number;
std::string data_type;
std::string ensemble_id;
std::string ensemble_label;
std::string creator;
std::string creator_hardware;
std::string creation_date;
std::string archive_date;
std::string floating_point;
};
}
#else
namespace Grid {
GRID_SERIALIZABLE_CLASS_MEMBERS(scidacFile,
double, version,
int, spacetime,
std::string, dims, // must convert to int
int, volfmt);
struct ILDGtype {
bool is_ILDG;
ILDGtype() : is_ILDG(false) {}
};
}
std::vector<int> getDimensions(void) {
std::stringstream stream(dims);
std::vector<int> dimensions;
int n;
while(stream >> n){
dimensions.push_back(n);
}
return dimensions;
}
void setDimensions(std::vector<int> dimensions) {
char delimiter = ' ';
std::stringstream stream;
for(int i=0;i<dimensions.size();i++){
stream << dimensions[i];
if ( i != dimensions.size()-1) {
stream << delimiter <<std::endl;
}
}
dims = stream.str();
}
// Constructor provides Grid
scidacFile() =default; // default constructor
scidacFile(GridBase * grid){
version = 1.0;
spacetime = grid->_ndimension;
setDimensions(grid->FullDimensions());
volfmt = GRID_IO_SINGLEFILE;
}
};
///////////////////////////////////////////////////////////////////////
// scidac-private-record-xml : example
// <scidacRecord>
// <version>1.1</version><date>Tue Jul 26 21:14:44 2011 UTC</date><recordtype>0</recordtype>
// <datatype>QDP_D3_ColorMatrix</datatype><precision>D</precision><colors>3</colors><spins>4</spins>
// <typesize>144</typesize><datacount>4</datacount>
// </scidacRecord>
///////////////////////////////////////////////////////////////////////
struct scidacRecord : Serializable {
public:
GRID_SERIALIZABLE_CLASS_MEMBERS(scidacRecord,
double, version,
std::string, date,
int, recordtype,
std::string, datatype,
std::string, precision,
int, colors,
int, spins,
int, typesize,
int, datacount);
scidacRecord() { version =1.0; }
};
////////////////////////
// ILDG format
////////////////////////
struct ildgFormat : Serializable {
public:
GRID_SERIALIZABLE_CLASS_MEMBERS(ildgFormat,
double, version,
std::string, field,
int, precision,
int, lx,
int, ly,
int, lz,
int, lt);
ildgFormat() { version=1.0; };
};
////////////////////////
// USQCD info
////////////////////////
struct usqcdInfo : Serializable {
public:
GRID_SERIALIZABLE_CLASS_MEMBERS(usqcdInfo,
double, version,
double, plaq,
double, linktr,
std::string, info);
usqcdInfo() {
version=1.0;
};
};
////////////////////////
// Scidac Checksum
////////////////////////
struct scidacChecksum : Serializable {
public:
GRID_SERIALIZABLE_CLASS_MEMBERS(scidacChecksum,
double, version,
std::string, suma,
std::string, sumb);
scidacChecksum() {
version=1.0;
};
};
////////////////////////////////////////////////////////////////////////////////////////////////////////////////
// Type: scidac-file-xml <title>MILC ILDG archival gauge configuration</title>
////////////////////////////////////////////////////////////////////////////////////////////////////////////////
////////////////////////////////////////////////////////////////////////////////////////////////////////////////
// Type:
////////////////////////////////////////////////////////////////////////////////////////////////////////////////
////////////////////////
// Scidac private file xml
// <?xml version="1.0" encoding="UTF-8"?><scidacFile><version>1.1</version><spacetime>4</spacetime><dims>16 16 16 32 </dims><volfmt>0</volfmt></scidacFile>
////////////////////////
#if 0
////////////////////////////////////////////////////////////////////////////////////////
// From http://www.physics.utah.edu/~detar/scidac/qio_2p3.pdf
////////////////////////////////////////////////////////////////////////////////////////
struct usqcdPropFile : Serializable {
public:
GRID_SERIALIZABLE_CLASS_MEMBERS(usqcdPropFile,
double, version,
std::string, type,
std::string, info);
usqcdPropFile() {
version=1.0;
};
};
struct usqcdSourceInfo : Serializable {
public:
GRID_SERIALIZABLE_CLASS_MEMBERS(usqcdSourceInfo,
double, version,
std::string, info);
usqcdSourceInfo() {
version=1.0;
};
};
struct usqcdPropInfo : Serializable {
public:
GRID_SERIALIZABLE_CLASS_MEMBERS(usqcdPropInfo,
double, version,
int, spin,
int, color,
std::string, info);
usqcdPropInfo() {
version=1.0;
};
};
#endif
}
#endif
#endif

325
lib/parallelIO/MetaData.h Normal file
View File

@ -0,0 +1,325 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/parallelIO/NerscIO.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 */
#include <algorithm>
#include <iostream>
#include <iomanip>
#include <fstream>
#include <map>
#include <unistd.h>
#include <sys/utsname.h>
#include <pwd.h>
namespace Grid {
///////////////////////////////////////////////////////
// Precision mapping
///////////////////////////////////////////////////////
template<class vobj> static std::string getFormatString (void)
{
std::string format;
typedef typename getPrecision<vobj>::real_scalar_type stype;
if ( sizeof(stype) == sizeof(float) ) {
format = std::string("IEEE32BIG");
}
if ( sizeof(stype) == sizeof(double) ) {
format = std::string("IEEE64BIG");
}
return format;
}
////////////////////////////////////////////////////////////////////////////////
// header specification/interpretation
////////////////////////////////////////////////////////////////////////////////
class FieldMetaData : Serializable {
public:
GRID_SERIALIZABLE_CLASS_MEMBERS(FieldMetaData,
int, nd,
std::vector<int>, dimension,
std::vector<std::string>, boundary,
int, data_start,
std::string, hdr_version,
std::string, storage_format,
double, link_trace,
double, plaquette,
uint32_t, checksum,
uint32_t, scidac_checksuma,
uint32_t, scidac_checksumb,
unsigned int, sequence_number,
std::string, data_type,
std::string, ensemble_id,
std::string, ensemble_label,
std::string, ildg_lfn,
std::string, creator,
std::string, creator_hardware,
std::string, creation_date,
std::string, archive_date,
std::string, floating_point);
FieldMetaData(void) {
nd=4;
dimension.resize(4);
boundary.resize(4);
}
};
namespace QCD {
using namespace Grid;
//////////////////////////////////////////////////////////////////////
// Bit and Physical Checksumming and QA of data
//////////////////////////////////////////////////////////////////////
inline void GridMetaData(GridBase *grid,FieldMetaData &header)
{
int nd = grid->_ndimension;
header.nd = nd;
header.dimension.resize(nd);
header.boundary.resize(nd);
for(int d=0;d<nd;d++) {
header.dimension[d] = grid->_fdimensions[d];
}
for(int d=0;d<nd;d++) {
header.boundary[d] = std::string("PERIODIC");
}
}
inline void MachineCharacteristics(FieldMetaData &header)
{
// Who
struct passwd *pw = getpwuid (getuid());
if (pw) header.creator = std::string(pw->pw_name);
// When
std::time_t t = std::time(nullptr);
std::tm tm_ = *std::localtime(&t);
std::ostringstream oss;
// oss << std::put_time(&tm_, "%c %Z");
header.creation_date = oss.str();
header.archive_date = header.creation_date;
// What
struct utsname name; uname(&name);
header.creator_hardware = std::string(name.nodename)+"-";
header.creator_hardware+= std::string(name.machine)+"-";
header.creator_hardware+= std::string(name.sysname)+"-";
header.creator_hardware+= std::string(name.release);
}
#define dump_meta_data(field, s) \
s << "BEGIN_HEADER" << std::endl; \
s << "HDR_VERSION = " << field.hdr_version << std::endl; \
s << "DATATYPE = " << field.data_type << std::endl; \
s << "STORAGE_FORMAT = " << field.storage_format << std::endl; \
for(int i=0;i<4;i++){ \
s << "DIMENSION_" << i+1 << " = " << field.dimension[i] << std::endl ; \
} \
s << "LINK_TRACE = " << std::setprecision(10) << field.link_trace << std::endl; \
s << "PLAQUETTE = " << std::setprecision(10) << field.plaquette << std::endl; \
for(int i=0;i<4;i++){ \
s << "BOUNDARY_"<<i+1<<" = " << field.boundary[i] << std::endl; \
} \
\
s << "CHECKSUM = "<< std::hex << std::setw(10) << field.checksum << std::dec<<std::endl; \
s << "SCIDAC_CHECKSUMA = "<< std::hex << std::setw(10) << field.scidac_checksuma << std::dec<<std::endl; \
s << "SCIDAC_CHECKSUMB = "<< std::hex << std::setw(10) << field.scidac_checksumb << std::dec<<std::endl; \
s << "ENSEMBLE_ID = " << field.ensemble_id << std::endl; \
s << "ENSEMBLE_LABEL = " << field.ensemble_label << std::endl; \
s << "SEQUENCE_NUMBER = " << field.sequence_number << std::endl; \
s << "CREATOR = " << field.creator << std::endl; \
s << "CREATOR_HARDWARE = "<< field.creator_hardware << std::endl; \
s << "CREATION_DATE = " << field.creation_date << std::endl; \
s << "ARCHIVE_DATE = " << field.archive_date << std::endl; \
s << "FLOATING_POINT = " << field.floating_point << std::endl; \
s << "END_HEADER" << std::endl;
template<class vobj> inline void PrepareMetaData(Lattice<vobj> & field, FieldMetaData &header)
{
GridBase *grid = field._grid;
std::string format = getFormatString<vobj>();
header.floating_point = format;
header.checksum = 0x0; // Nersc checksum unused in ILDG, Scidac
GridMetaData(grid,header);
MachineCharacteristics(header);
}
inline void GaugeStatistics(Lattice<vLorentzColourMatrixF> & data,FieldMetaData &header)
{
// How to convert data precision etc...
header.link_trace=Grid::QCD::WilsonLoops<PeriodicGimplF>::linkTrace(data);
header.plaquette =Grid::QCD::WilsonLoops<PeriodicGimplF>::avgPlaquette(data);
}
inline void GaugeStatistics(Lattice<vLorentzColourMatrixD> & data,FieldMetaData &header)
{
// How to convert data precision etc...
header.link_trace=Grid::QCD::WilsonLoops<PeriodicGimplD>::linkTrace(data);
header.plaquette =Grid::QCD::WilsonLoops<PeriodicGimplD>::avgPlaquette(data);
}
template<> inline void PrepareMetaData<vLorentzColourMatrixF>(Lattice<vLorentzColourMatrixF> & field, FieldMetaData &header)
{
GridBase *grid = field._grid;
std::string format = getFormatString<vLorentzColourMatrixF>();
header.floating_point = format;
header.checksum = 0x0; // Nersc checksum unused in ILDG, Scidac
GridMetaData(grid,header);
GaugeStatistics(field,header);
MachineCharacteristics(header);
}
template<> inline void PrepareMetaData<vLorentzColourMatrixD>(Lattice<vLorentzColourMatrixD> & field, FieldMetaData &header)
{
GridBase *grid = field._grid;
std::string format = getFormatString<vLorentzColourMatrixD>();
header.floating_point = format;
header.checksum = 0x0; // Nersc checksum unused in ILDG, Scidac
GridMetaData(grid,header);
GaugeStatistics(field,header);
MachineCharacteristics(header);
}
//////////////////////////////////////////////////////////////////////
// Utilities ; these are QCD aware
//////////////////////////////////////////////////////////////////////
inline void reconstruct3(LorentzColourMatrix & cm)
{
const int x=0;
const int y=1;
const int z=2;
for(int mu=0;mu<Nd;mu++){
cm(mu)()(2,x) = adj(cm(mu)()(0,y)*cm(mu)()(1,z)-cm(mu)()(0,z)*cm(mu)()(1,y)); //x= yz-zy
cm(mu)()(2,y) = adj(cm(mu)()(0,z)*cm(mu)()(1,x)-cm(mu)()(0,x)*cm(mu)()(1,z)); //y= zx-xz
cm(mu)()(2,z) = adj(cm(mu)()(0,x)*cm(mu)()(1,y)-cm(mu)()(0,y)*cm(mu)()(1,x)); //z= xy-yx
}
}
////////////////////////////////////////////////////////////////////////////////
// Some data types for intermediate storage
////////////////////////////////////////////////////////////////////////////////
template<typename vtype> using iLorentzColour2x3 = iVector<iVector<iVector<vtype, Nc>, 2>, Nd >;
typedef iLorentzColour2x3<Complex> LorentzColour2x3;
typedef iLorentzColour2x3<ComplexF> LorentzColour2x3F;
typedef iLorentzColour2x3<ComplexD> LorentzColour2x3D;
/////////////////////////////////////////////////////////////////////////////////
// Simple classes for precision conversion
/////////////////////////////////////////////////////////////////////////////////
template <class fobj, class sobj>
struct BinarySimpleUnmunger {
typedef typename getPrecision<fobj>::real_scalar_type fobj_stype;
typedef typename getPrecision<sobj>::real_scalar_type sobj_stype;
void operator()(sobj &in, fobj &out) {
// take word by word and transform accoding to the status
fobj_stype *out_buffer = (fobj_stype *)&out;
sobj_stype *in_buffer = (sobj_stype *)&in;
size_t fobj_words = sizeof(out) / sizeof(fobj_stype);
size_t sobj_words = sizeof(in) / sizeof(sobj_stype);
assert(fobj_words == sobj_words);
for (unsigned int word = 0; word < sobj_words; word++)
out_buffer[word] = in_buffer[word]; // type conversion on the fly
}
};
template <class fobj, class sobj>
struct BinarySimpleMunger {
typedef typename getPrecision<fobj>::real_scalar_type fobj_stype;
typedef typename getPrecision<sobj>::real_scalar_type sobj_stype;
void operator()(fobj &in, sobj &out) {
// take word by word and transform accoding to the status
fobj_stype *in_buffer = (fobj_stype *)&in;
sobj_stype *out_buffer = (sobj_stype *)&out;
size_t fobj_words = sizeof(in) / sizeof(fobj_stype);
size_t sobj_words = sizeof(out) / sizeof(sobj_stype);
assert(fobj_words == sobj_words);
for (unsigned int word = 0; word < sobj_words; word++)
out_buffer[word] = in_buffer[word]; // type conversion on the fly
}
};
template<class fobj,class sobj>
struct GaugeSimpleMunger{
void operator()(fobj &in, sobj &out) {
for (int mu = 0; mu < Nd; mu++) {
for (int i = 0; i < Nc; i++) {
for (int j = 0; j < Nc; j++) {
out(mu)()(i, j) = in(mu)()(i, j);
}}
}
};
};
template <class fobj, class sobj>
struct GaugeSimpleUnmunger {
void operator()(sobj &in, fobj &out) {
for (int mu = 0; mu < Nd; mu++) {
for (int i = 0; i < Nc; i++) {
for (int j = 0; j < Nc; j++) {
out(mu)()(i, j) = in(mu)()(i, j);
}}
}
};
};
template<class fobj,class sobj>
struct Gauge3x2munger{
void operator() (fobj &in,sobj &out){
for(int mu=0;mu<Nd;mu++){
for(int i=0;i<2;i++){
for(int j=0;j<3;j++){
out(mu)()(i,j) = in(mu)(i)(j);
}}
}
reconstruct3(out);
}
};
template<class fobj,class sobj>
struct Gauge3x2unmunger{
void operator() (sobj &in,fobj &out){
for(int mu=0;mu<Nd;mu++){
for(int i=0;i<2;i++){
for(int j=0;j<3;j++){
out(mu)(i)(j) = in(mu)()(i,j);
}}
}
}
};
}
}

383
lib/parallelIO/NerscIO.h
View File

@ -30,182 +30,11 @@
#ifndef GRID_NERSC_IO_H
#define GRID_NERSC_IO_H
#include <algorithm>
#include <iostream>
#include <iomanip>
#include <fstream>
#include <map>
#include <unistd.h>
#include <sys/utsname.h>
#include <pwd.h>
namespace Grid {
namespace QCD {
using namespace Grid;
////////////////////////////////////////////////////////////////////////////////
// Some data types for intermediate storage
////////////////////////////////////////////////////////////////////////////////
template<typename vtype> using iLorentzColour2x3 = iVector<iVector<iVector<vtype, Nc>, 2>, 4 >;
typedef iLorentzColour2x3<Complex> LorentzColour2x3;
typedef iLorentzColour2x3<ComplexF> LorentzColour2x3F;
typedef iLorentzColour2x3<ComplexD> LorentzColour2x3D;
////////////////////////////////////////////////////////////////////////////////
// header specification/interpretation
////////////////////////////////////////////////////////////////////////////////
class NerscField {
public:
// header strings (not in order)
int dimension[4];
std::string boundary[4];
int data_start;
std::string hdr_version;
std::string storage_format;
// Checks on data
double link_trace;
double plaquette;
uint32_t checksum;
unsigned int sequence_number;
std::string data_type;
std::string ensemble_id ;
std::string ensemble_label ;
std::string creator ;
std::string creator_hardware ;
std::string creation_date ;
std::string archive_date ;
std::string floating_point;
};
//////////////////////////////////////////////////////////////////////
// Bit and Physical Checksumming and QA of data
//////////////////////////////////////////////////////////////////////
inline void NerscGrid(GridBase *grid,NerscField &header)
{
assert(grid->_ndimension==4);
for(int d=0;d<4;d++) {
header.dimension[d] = grid->_fdimensions[d];
}
for(int d=0;d<4;d++) {
header.boundary[d] = std::string("PERIODIC");
}
}
template<class GaugeField>
inline void NerscStatistics(GaugeField & data,NerscField &header)
{
// How to convert data precision etc...
header.link_trace=Grid::QCD::WilsonLoops<PeriodicGimplR>::linkTrace(data);
header.plaquette =Grid::QCD::WilsonLoops<PeriodicGimplR>::avgPlaquette(data);
}
inline void NerscMachineCharacteristics(NerscField &header)
{
// Who
struct passwd *pw = getpwuid (getuid());
if (pw) header.creator = std::string(pw->pw_name);
// When
std::time_t t = std::time(nullptr);
std::tm tm = *std::localtime(&t);
std::ostringstream oss;
// oss << std::put_time(&tm, "%c %Z");
header.creation_date = oss.str();
header.archive_date = header.creation_date;
// What
struct utsname name; uname(&name);
header.creator_hardware = std::string(name.nodename)+"-";
header.creator_hardware+= std::string(name.machine)+"-";
header.creator_hardware+= std::string(name.sysname)+"-";
header.creator_hardware+= std::string(name.release);
}
//////////////////////////////////////////////////////////////////////
// Utilities ; these are QCD aware
//////////////////////////////////////////////////////////////////////
inline void NerscChecksum(uint32_t *buf,uint32_t buf_size_bytes,uint32_t &csum)
{
BinaryIO::Uint32Checksum(buf,buf_size_bytes,csum);
}
inline void reconstruct3(LorentzColourMatrix & cm)
{
const int x=0;
const int y=1;
const int z=2;
for(int mu=0;mu<4;mu++){
cm(mu)()(2,x) = adj(cm(mu)()(0,y)*cm(mu)()(1,z)-cm(mu)()(0,z)*cm(mu)()(1,y)); //x= yz-zy
cm(mu)()(2,y) = adj(cm(mu)()(0,z)*cm(mu)()(1,x)-cm(mu)()(0,x)*cm(mu)()(1,z)); //y= zx-xz
cm(mu)()(2,z) = adj(cm(mu)()(0,x)*cm(mu)()(1,y)-cm(mu)()(0,y)*cm(mu)()(1,x)); //z= xy-yx
}
}
template<class fobj,class sobj>
struct NerscSimpleMunger{
void operator()(fobj &in, sobj &out, uint32_t &csum) {
for (int mu = 0; mu < Nd; mu++) {
for (int i = 0; i < Nc; i++) {
for (int j = 0; j < Nc; j++) {
out(mu)()(i, j) = in(mu)()(i, j);
}
}
}
NerscChecksum((uint32_t *)&in, sizeof(in), csum);
};
};
template <class fobj, class sobj>
struct NerscSimpleUnmunger {
void operator()(sobj &in, fobj &out, uint32_t &csum) {
for (int mu = 0; mu < Nd; mu++) {
for (int i = 0; i < Nc; i++) {
for (int j = 0; j < Nc; j++) {
out(mu)()(i, j) = in(mu)()(i, j);
}
}
}
NerscChecksum((uint32_t *)&out, sizeof(out), csum);
};
};
template<class fobj,class sobj>
struct Nersc3x2munger{
void operator() (fobj &in,sobj &out,uint32_t &csum){
NerscChecksum((uint32_t *)&in,sizeof(in),csum);
for(int mu=0;mu<4;mu++){
for(int i=0;i<2;i++){
for(int j=0;j<3;j++){
out(mu)()(i,j) = in(mu)(i)(j);
}}
}
reconstruct3(out);
}
};
template<class fobj,class sobj>
struct Nersc3x2unmunger{
void operator() (sobj &in,fobj &out,uint32_t &csum){
for(int mu=0;mu<4;mu++){
for(int i=0;i<2;i++){
for(int j=0;j<3;j++){
out(mu)(i)(j) = in(mu)()(i,j);
}}
}
NerscChecksum((uint32_t *)&out,sizeof(out),csum);
}
};
////////////////////////////////////////////////////////////////////////////////
// Write and read from fstream; comput header offset for payload
////////////////////////////////////////////////////////////////////////////////
@ -216,42 +45,17 @@ namespace Grid {
std::ofstream fout(file,std::ios::out);
}
#define dump_nersc_header(field, s) \
s << "BEGIN_HEADER" << std::endl; \
s << "HDR_VERSION = " << field.hdr_version << std::endl; \
s << "DATATYPE = " << field.data_type << std::endl; \
s << "STORAGE_FORMAT = " << field.storage_format << std::endl; \
for(int i=0;i<4;i++){ \
s << "DIMENSION_" << i+1 << " = " << field.dimension[i] << std::endl ; \
} \
s << "LINK_TRACE = " << std::setprecision(10) << field.link_trace << std::endl; \
s << "PLAQUETTE = " << std::setprecision(10) << field.plaquette << std::endl; \
for(int i=0;i<4;i++){ \
s << "BOUNDARY_"<<i+1<<" = " << field.boundary[i] << std::endl; \
} \
\
s << "CHECKSUM = "<< std::hex << std::setw(10) << field.checksum << std::dec<<std::endl; \
s << "ENSEMBLE_ID = " << field.ensemble_id << std::endl; \
s << "ENSEMBLE_LABEL = " << field.ensemble_label << std::endl; \
s << "SEQUENCE_NUMBER = " << field.sequence_number << std::endl; \
s << "CREATOR = " << field.creator << std::endl; \
s << "CREATOR_HARDWARE = "<< field.creator_hardware << std::endl; \
s << "CREATION_DATE = " << field.creation_date << std::endl; \
s << "ARCHIVE_DATE = " << field.archive_date << std::endl; \
s << "FLOATING_POINT = " << field.floating_point << std::endl; \
s << "END_HEADER" << std::endl;
static inline unsigned int writeHeader(NerscField &field,std::string file)
static inline unsigned int writeHeader(FieldMetaData &field,std::string file)
{
std::ofstream fout(file,std::ios::out|std::ios::in);
fout.seekp(0,std::ios::beg);
dump_nersc_header(field, fout);
dump_meta_data(field, fout);
field.data_start = fout.tellp();
return field.data_start;
}
// for the header-reader
static inline int readHeader(std::string file,GridBase *grid, NerscField &field)
static inline int readHeader(std::string file,GridBase *grid, FieldMetaData &field)
{
int offset=0;
std::map<std::string,std::string> header;
@ -323,21 +127,21 @@ namespace Grid {
return field.data_start;
}
/////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
// Now the meat: the object readers
/////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
#define PARALLEL_READ
#define PARALLEL_WRITE
/////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
// Now the meat: the object readers
/////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
template<class vsimd>
static inline void readConfiguration(Lattice<iLorentzColourMatrix<vsimd> > &Umu,NerscField& header,std::string file)
{
template<class vsimd>
static inline void readConfiguration(Lattice<iLorentzColourMatrix<vsimd> > &Umu,
FieldMetaData& header,
std::string file)
{
typedef Lattice<iLorentzColourMatrix<vsimd> > GaugeField;
GridBase *grid = Umu._grid;
int offset = readHeader(file,Umu._grid,header);
NerscField clone(header);
FieldMetaData clone(header);
std::string format(header.floating_point);
@ -346,76 +150,78 @@ namespace Grid {
int ieee64big = (format == std::string("IEEE64BIG"));
int ieee64 = (format == std::string("IEEE64"));
uint32_t csum;
uint32_t nersc_csum,scidac_csuma,scidac_csumb;
// depending on datatype, set up munger;
// munger is a function of <floating point, Real, data_type>
if ( header.data_type == std::string("4D_SU3_GAUGE") ) {
if ( ieee32 || ieee32big ) {
#ifdef PARALLEL_READ
csum=BinaryIO::readObjectParallel<iLorentzColourMatrix<vsimd>, LorentzColour2x3F>
(Umu,file,Nersc3x2munger<LorentzColour2x3F,LorentzColourMatrix>(), offset,format);
#else
csum=BinaryIO::readObjectSerial<iLorentzColourMatrix<vsimd>, LorentzColour2x3F>
(Umu,file,Nersc3x2munger<LorentzColour2x3F,LorentzColourMatrix>(), offset,format);
#endif
}
if ( ieee64 || ieee64big ) {
#ifdef PARALLEL_READ
csum=BinaryIO::readObjectParallel<iLorentzColourMatrix<vsimd>, LorentzColour2x3D>
(Umu,file,Nersc3x2munger<LorentzColour2x3D,LorentzColourMatrix>(),offset,format);
#else
csum=BinaryIO::readObjectSerial<iLorentzColourMatrix<vsimd>, LorentzColour2x3D>
(Umu,file,Nersc3x2munger<LorentzColour2x3D,LorentzColourMatrix>(),offset,format);
#endif
}
} else if ( header.data_type == std::string("4D_SU3_GAUGE_3x3") ) {
if ( ieee32 || ieee32big ) {
#ifdef PARALLEL_READ
csum=BinaryIO::readObjectParallel<iLorentzColourMatrix<vsimd>,LorentzColourMatrixF>
(Umu,file,NerscSimpleMunger<LorentzColourMatrixF,LorentzColourMatrix>(),offset,format);
#else
csum=BinaryIO::readObjectSerial<iLorentzColourMatrix<vsimd>,LorentzColourMatrixF>
(Umu,file,NerscSimpleMunger<LorentzColourMatrixF,LorentzColourMatrix>(),offset,format);
#endif
BinaryIO::readLatticeObject<iLorentzColourMatrix<vsimd>, LorentzColour2x3F>
(Umu,file,Gauge3x2munger<LorentzColour2x3F,LorentzColourMatrix>(), offset,format,
nersc_csum,scidac_csuma,scidac_csumb);
}
if ( ieee64 || ieee64big ) {
#ifdef PARALLEL_READ
csum=BinaryIO::readObjectParallel<iLorentzColourMatrix<vsimd>,LorentzColourMatrixD>
(Umu,file,NerscSimpleMunger<LorentzColourMatrixD,LorentzColourMatrix>(),offset,format);
#else
csum=BinaryIO::readObjectSerial<iLorentzColourMatrix<vsimd>,LorentzColourMatrixD>
(Umu,file,NerscSimpleMunger<LorentzColourMatrixD,LorentzColourMatrix>(),offset,format);
#endif
BinaryIO::readLatticeObject<iLorentzColourMatrix<vsimd>, LorentzColour2x3D>
(Umu,file,Gauge3x2munger<LorentzColour2x3D,LorentzColourMatrix>(),offset,format,
nersc_csum,scidac_csuma,scidac_csumb);
}
} else if ( header.data_type == std::string("4D_SU3_GAUGE_3x3") ) {
if ( ieee32 || ieee32big ) {
BinaryIO::readLatticeObject<iLorentzColourMatrix<vsimd>,LorentzColourMatrixF>
(Umu,file,GaugeSimpleMunger<LorentzColourMatrixF,LorentzColourMatrix>(),offset,format,
nersc_csum,scidac_csuma,scidac_csumb);
}
if ( ieee64 || ieee64big ) {
BinaryIO::readLatticeObject<iLorentzColourMatrix<vsimd>,LorentzColourMatrixD>
(Umu,file,GaugeSimpleMunger<LorentzColourMatrixD,LorentzColourMatrix>(),offset,format,
nersc_csum,scidac_csuma,scidac_csumb);
}
} else {
assert(0);
}
NerscStatistics<GaugeField>(Umu,clone);
GaugeStatistics(Umu,clone);
std::cout<<GridLogMessage <<"NERSC Configuration "<<file<<" checksum "<<std::hex<< csum<< std::dec
std::cout<<GridLogMessage <<"NERSC Configuration "<<file<<" checksum "<<std::hex<<nersc_csum<< std::dec
<<" header "<<std::hex<<header.checksum<<std::dec <<std::endl;
std::cout<<GridLogMessage <<"NERSC Configuration "<<file<<" plaquette "<<clone.plaquette
<<" header "<<header.plaquette<<std::endl;
std::cout<<GridLogMessage <<"NERSC Configuration "<<file<<" link_trace "<<clone.link_trace
<<" header "<<header.link_trace<<std::endl;
if ( fabs(clone.plaquette -header.plaquette ) >= 1.0e-5 ) {
std::cout << " Plaquette mismatch "<<std::endl;
std::cout << Umu[0]<<std::endl;
std::cout << Umu[1]<<std::endl;
}
if ( nersc_csum != header.checksum ) {
std::cerr << " checksum mismatch " << std::endl;
std::cerr << " plaqs " << clone.plaquette << " " << header.plaquette << std::endl;
std::cerr << " trace " << clone.link_trace<< " " << header.link_trace<< std::endl;
std::cerr << " nersc_csum " <<std::hex<< nersc_csum << " " << header.checksum<< std::dec<< std::endl;
exit(0);
}
assert(fabs(clone.plaquette -header.plaquette ) < 1.0e-5 );
assert(fabs(clone.link_trace-header.link_trace) < 1.0e-6 );
assert(csum == header.checksum );
assert(nersc_csum == header.checksum );
std::cout<<GridLogMessage <<"NERSC Configuration "<<file<< " and plaquette, link trace, and checksum agree"<<std::endl;
}
}
template<class vsimd>
static inline void writeConfiguration(Lattice<iLorentzColourMatrix<vsimd> > &Umu,std::string file, int two_row,int bits32)
static inline void writeConfiguration(Lattice<iLorentzColourMatrix<vsimd> > &Umu,
std::string file,
int two_row,
int bits32)
{
typedef Lattice<iLorentzColourMatrix<vsimd> > GaugeField;
typedef iLorentzColourMatrix<vsimd> vobj;
typedef typename vobj::scalar_object sobj;
FieldMetaData header;
///////////////////////////////////////////
// Following should become arguments
NerscField header;
///////////////////////////////////////////
header.sequence_number = 1;
header.ensemble_id = "UKQCD";
header.ensemble_label = "DWF";
@ -425,45 +231,32 @@ namespace Grid {
GridBase *grid = Umu._grid;
NerscGrid(grid,header);
NerscStatistics<GaugeField>(Umu,header);
NerscMachineCharacteristics(header);
GridMetaData(grid,header);
assert(header.nd==4);
GaugeStatistics(Umu,header);
MachineCharacteristics(header);
uint32_t csum;
int offset;
truncate(file);
if ( two_row ) {
// Sod it -- always write 3x3 double
header.floating_point = std::string("IEEE64BIG");
header.data_type = std::string("4D_SU3_GAUGE_3x3");
GaugeSimpleUnmunger<fobj3D,sobj> munge;
offset = writeHeader(header,file);
header.floating_point = std::string("IEEE64BIG");
header.data_type = std::string("4D_SU3_GAUGE");
Nersc3x2unmunger<fobj2D,sobj> munge;
BinaryIO::Uint32Checksum<vobj,fobj2D>(Umu, munge,header.checksum);
offset = writeHeader(header,file);
#ifdef PARALLEL_WRITE
csum=BinaryIO::writeObjectParallel<vobj,fobj2D>(Umu,file,munge,offset,header.floating_point);
#else
csum=BinaryIO::writeObjectSerial<vobj,fobj2D>(Umu,file,munge,offset,header.floating_point);
#endif
} else {
header.floating_point = std::string("IEEE64BIG");
header.data_type = std::string("4D_SU3_GAUGE_3x3");
NerscSimpleUnmunger<fobj3D,sobj> munge;
BinaryIO::Uint32Checksum<vobj,fobj3D>(Umu, munge,header.checksum);
offset = writeHeader(header,file);
#ifdef PARALLEL_WRITE
csum=BinaryIO::writeObjectParallel<vobj,fobj3D>(Umu,file,munge,offset,header.floating_point);
#else
csum=BinaryIO::writeObjectSerial<vobj,fobj3D>(Umu,file,munge,offset,header.floating_point);
#endif
}
uint32_t nersc_csum,scidac_csuma,scidac_csumb;
BinaryIO::writeLatticeObject<vobj,fobj3D>(Umu,file,munge,offset,header.floating_point,
nersc_csum,scidac_csuma,scidac_csumb);
header.checksum = nersc_csum;
writeHeader(header,file);
std::cout<<GridLogMessage <<"Written NERSC Configuration on "<< file << " checksum "<<std::hex<<csum<< std::dec<<" plaq "<< header.plaquette <<std::endl;
std::cout<<GridLogMessage <<"Written NERSC Configuration on "<< file << " checksum "
<<std::hex<<header.checksum
<<std::dec<<" plaq "<< header.plaquette <<std::endl;
}
///////////////////////////////
// RNG state
///////////////////////////////
@ -472,19 +265,19 @@ namespace Grid {
typedef typename GridParallelRNG::RngStateType RngStateType;
// Following should become arguments
NerscField header;
FieldMetaData header;
header.sequence_number = 1;
header.ensemble_id = "UKQCD";
header.ensemble_label = "DWF";
GridBase *grid = parallel._grid;
NerscGrid(grid,header);
GridMetaData(grid,header);
assert(header.nd==4);
header.link_trace=0.0;
header.plaquette=0.0;
NerscMachineCharacteristics(header);
MachineCharacteristics(header);
uint32_t csum;
int offset;
#ifdef RNG_RANLUX
@ -502,15 +295,19 @@ namespace Grid {
truncate(file);
offset = writeHeader(header,file);
csum=BinaryIO::writeRNGSerial(serial,parallel,file,offset);
header.checksum = csum;
uint32_t nersc_csum,scidac_csuma,scidac_csumb;
BinaryIO::writeRNG(serial,parallel,file,offset,nersc_csum,scidac_csuma,scidac_csumb);
header.checksum = nersc_csum;
offset = writeHeader(header,file);
std::cout<<GridLogMessage <<"Written NERSC RNG STATE "<<file<< " checksum "<<std::hex<<csum<<std::dec<<std::endl;
std::cout<<GridLogMessage
<<"Written NERSC RNG STATE "<<file<< " checksum "
<<std::hex<<header.checksum
<<std::dec<<std::endl;
}
static inline void readRNGState(GridSerialRNG &serial,GridParallelRNG & parallel,NerscField& header,std::string file)
static inline void readRNGState(GridSerialRNG &serial,GridParallelRNG & parallel,FieldMetaData& header,std::string file)
{
typedef typename GridParallelRNG::RngStateType RngStateType;
@ -518,7 +315,7 @@ namespace Grid {
int offset = readHeader(file,grid,header);
NerscField clone(header);
FieldMetaData clone(header);
std::string format(header.floating_point);
std::string data_type(header.data_type);
@ -538,15 +335,19 @@ namespace Grid {
// depending on datatype, set up munger;
// munger is a function of <floating point, Real, data_type>
uint32_t csum=BinaryIO::readRNGSerial(serial,parallel,file,offset);
uint32_t nersc_csum,scidac_csuma,scidac_csumb;
BinaryIO::readRNG(serial,parallel,file,offset,nersc_csum,scidac_csuma,scidac_csumb);
assert(csum == header.checksum );
if ( nersc_csum != header.checksum ) {
std::cerr << "checksum mismatch "<<std::hex<< nersc_csum <<" "<<header.checksum<<std::dec<<std::endl;
exit(0);
}
assert(nersc_csum == header.checksum );
std::cout<<GridLogMessage <<"Read NERSC RNG file "<<file<< " format "<< data_type <<std::endl;
}
};
}}
#endif

2
lib/perfmon/PerfCount.cc
View File

@ -40,7 +40,7 @@ const PerformanceCounter::PerformanceCounterConfig PerformanceCounter::Performan
{ PERF_TYPE_HARDWARE, PERF_COUNT_HW_CPU_CYCLES , "CPUCYCLES.........." , INSTRUCTIONS},
{ PERF_TYPE_HARDWARE, PERF_COUNT_HW_INSTRUCTIONS , "INSTRUCTIONS......." , CPUCYCLES },
// 4
#ifdef AVX512
#ifdef KNL
{ PERF_TYPE_RAW, RawConfig(0x40,0x04), "ALL_LOADS..........", CPUCYCLES },
{ PERF_TYPE_RAW, RawConfig(0x01,0x04), "L1_MISS_LOADS......", L1D_READ_ACCESS },
{ PERF_TYPE_RAW, RawConfig(0x40,0x04), "ALL_LOADS..........", L1D_READ_ACCESS },

14
lib/qcd/action/fermion/CayleyFermion5D.cc
View File

@ -414,7 +414,7 @@ void CayleyFermion5D<Impl>::SetCoefficientsInternal(RealD zolo_hi,std::vector<Co
for(int i=0; i < Ls; i++){
as[i] = 1.0;
omega[i] = gamma[i]*zolo_hi; //NB reciprocal relative to Chroma NEF code
// assert(fabs(omega[i])>0.0);
assert(omega[i]!=Coeff_t(0.0));
bs[i] = 0.5*(bpc/omega[i] + bmc);
cs[i] = 0.5*(bpc/omega[i] - bmc);
}
@ -429,7 +429,7 @@ void CayleyFermion5D<Impl>::SetCoefficientsInternal(RealD zolo_hi,std::vector<Co
for(int i=0;i<Ls;i++){
bee[i]=as[i]*(bs[i]*(4.0-this->M5) +1.0);
// assert(fabs(bee[i])>0.0);
assert(bee[i]!=Coeff_t(0.0));
cee[i]=as[i]*(1.0-cs[i]*(4.0-this->M5));
beo[i]=as[i]*bs[i];
ceo[i]=-as[i]*cs[i];
@ -455,11 +455,17 @@ void CayleyFermion5D<Impl>::SetCoefficientsInternal(RealD zolo_hi,std::vector<Co
dee[i] = bee[i];
if ( i < Ls-1 ) {
assert(bee[i]!=Coeff_t(0.0));
assert(bee[0]!=Coeff_t(0.0));
lee[i] =-cee[i+1]/bee[i]; // sub-diag entry on the ith column
leem[i]=mass*cee[Ls-1]/bee[0];
for(int j=0;j<i;j++) leem[i]*= aee[j]/bee[j+1];
for(int j=0;j<i;j++) {
assert(bee[j+1]!=Coeff_t(0.0));
leem[i]*= aee[j]/bee[j+1];
}
uee[i] =-aee[i]/bee[i]; // up-diag entry on the ith row
@ -478,7 +484,7 @@ void CayleyFermion5D<Impl>::SetCoefficientsInternal(RealD zolo_hi,std::vector<Co
{
Coeff_t delta_d=mass*cee[Ls-1];
for(int j=0;j<Ls-1;j++) {
// assert(fabs(bee[j])>0.0);
assert(bee[j] != Coeff_t(0.0));
delta_d *= cee[j]/bee[j];
}
dee[Ls-1] += delta_d;

7
lib/qcd/action/fermion/Fermion.h
View File

@ -237,4 +237,11 @@ typedef ImprovedStaggeredFermion5D<StaggeredVec5dImplD> ImprovedStaggeredFermion
}}
////////////////////
// Scalar QED actions
// TODO: this needs to move to another header after rename to Fermion.h
////////////////////
#include <Grid/qcd/action/scalar/Scalar.h>
#include <Grid/qcd/action/gauge/Photon.h>
#endif

6
lib/qcd/action/fermion/FermionOperatorImpl.h
View File

@ -644,19 +644,16 @@ class StaggeredImpl : public PeriodicGaugeImpl<GaugeImplTypes<S, Representation:
INHERIT_GIMPL_TYPES(Gimpl);
template <typename vtype> using iImplScalar = iScalar<iScalar<iScalar<vtype> > >;
template <typename vtype> using iImplSpinor = iScalar<iScalar<iVector<vtype, Dimension> > >;
template <typename vtype> using iImplHalfSpinor = iScalar<iScalar<iVector<vtype, Dimension> > >;
template <typename vtype> using iImplDoubledGaugeField = iVector<iScalar<iMatrix<vtype, Dimension> >, Nds>;
template <typename vtype> using iImplPropagator = iScalar<iScalar<iMatrix<vtype, Dimension> > >;
typedef iImplScalar<Simd> SiteComplex;
typedef iImplSpinor<Simd> SiteSpinor;
typedef iImplHalfSpinor<Simd> SiteHalfSpinor;
typedef iImplDoubledGaugeField<Simd> SiteDoubledGaugeField;
typedef iImplPropagator<Simd> SitePropagator;
typedef Lattice<SiteComplex> ComplexField;
typedef Lattice<SiteSpinor> FermionField;
typedef Lattice<SiteDoubledGaugeField> DoubledGaugeField;
typedef Lattice<SitePropagator> PropagatorField;
@ -775,7 +772,6 @@ class StaggeredImpl : public PeriodicGaugeImpl<GaugeImplTypes<S, Representation:
INHERIT_GIMPL_TYPES(Gimpl);
template <typename vtype> using iImplScalar = iScalar<iScalar<iScalar<vtype> > >;
template <typename vtype> using iImplSpinor = iScalar<iScalar<iVector<vtype, Dimension> > >;
template <typename vtype> using iImplHalfSpinor = iScalar<iScalar<iVector<vtype, Dimension> > >;
template <typename vtype> using iImplDoubledGaugeField = iVector<iScalar<iMatrix<vtype, Dimension> >, Nds>;
@ -792,12 +788,10 @@ class StaggeredImpl : public PeriodicGaugeImpl<GaugeImplTypes<S, Representation:
typedef Lattice<SiteDoubledGaugeField> DoubledGaugeField;
typedef Lattice<SitePropagator> PropagatorField;
typedef iImplScalar<Simd> SiteComplex;
typedef iImplSpinor<Simd> SiteSpinor;
typedef iImplHalfSpinor<Simd> SiteHalfSpinor;
typedef Lattice<SiteComplex> ComplexField;
typedef Lattice<SiteSpinor> FermionField;
typedef SimpleCompressor<SiteSpinor> Compressor;

60
lib/qcd/action/fermion/ImprovedStaggeredFermion5D.cc
View File

@ -230,8 +230,15 @@ void ImprovedStaggeredFermion5D<Impl>::DhopInternal(StencilImpl & st, LebesgueOr
{
Compressor compressor;
int LLs = in._grid->_rdimensions[0];
DhopTotalTime -= usecond();
DhopCommTime -= usecond();
st.HaloExchange(in,compressor);
DhopCommTime += usecond();
DhopComputeTime -= usecond();
// Dhop takes the 4d grid from U, and makes a 5d index for fermion
if (dag == DaggerYes) {
parallel_for (int ss = 0; ss < U._grid->oSites(); ss++) {
@ -244,12 +251,15 @@ void ImprovedStaggeredFermion5D<Impl>::DhopInternal(StencilImpl & st, LebesgueOr
Kernels::DhopSite(st,lo,U,UUU,st.CommBuf(),LLs,sU,in,out);
}
}
DhopComputeTime += usecond();
DhopTotalTime += usecond();
}
template<class Impl>
void ImprovedStaggeredFermion5D<Impl>::DhopOE(const FermionField &in, FermionField &out,int dag)
{
DhopCalls+=1;
conformable(in._grid,FermionRedBlackGrid()); // verifies half grid
conformable(in._grid,out._grid); // drops the cb check
@ -261,6 +271,7 @@ void ImprovedStaggeredFermion5D<Impl>::DhopOE(const FermionField &in, FermionFie
template<class Impl>
void ImprovedStaggeredFermion5D<Impl>::DhopEO(const FermionField &in, FermionField &out,int dag)
{
DhopCalls+=1;
conformable(in._grid,FermionRedBlackGrid()); // verifies half grid
conformable(in._grid,out._grid); // drops the cb check
@ -272,6 +283,7 @@ void ImprovedStaggeredFermion5D<Impl>::DhopEO(const FermionField &in, FermionFie
template<class Impl>
void ImprovedStaggeredFermion5D<Impl>::Dhop(const FermionField &in, FermionField &out,int dag)
{
DhopCalls+=2;
conformable(in._grid,FermionGrid()); // verifies full grid
conformable(in._grid,out._grid);
@ -280,6 +292,54 @@ void ImprovedStaggeredFermion5D<Impl>::Dhop(const FermionField &in, FermionField
DhopInternal(Stencil,Lebesgue,Umu,UUUmu,in,out,dag);
}
template<class Impl>
void ImprovedStaggeredFermion5D<Impl>::Report(void)
{
std::vector<int> latt = GridDefaultLatt();
RealD volume = Ls; for(int mu=0;mu<Nd;mu++) volume=volume*latt[mu];
RealD NP = _FourDimGrid->_Nprocessors;
RealD NN = _FourDimGrid->NodeCount();
std::cout << GridLogMessage << "#### Dhop calls report " << std::endl;
std::cout << GridLogMessage << "ImprovedStaggeredFermion5D Number of DhopEO Calls : "
<< DhopCalls << std::endl;
std::cout << GridLogMessage << "ImprovedStaggeredFermion5D TotalTime /Calls : "
<< DhopTotalTime / DhopCalls << " us" << std::endl;
std::cout << GridLogMessage << "ImprovedStaggeredFermion5D CommTime /Calls : "
<< DhopCommTime / DhopCalls << " us" << std::endl;
std::cout << GridLogMessage << "ImprovedStaggeredFermion5D ComputeTime/Calls : "
<< DhopComputeTime / DhopCalls << " us" << std::endl;
// Average the compute time
_FourDimGrid->GlobalSum(DhopComputeTime);
DhopComputeTime/=NP;
RealD mflops = 1154*volume*DhopCalls/DhopComputeTime/2; // 2 for red black counting
std::cout << GridLogMessage << "Average mflops/s per call : " << mflops << std::endl;
std::cout << GridLogMessage << "Average mflops/s per call per rank : " << mflops/NP << std::endl;
std::cout << GridLogMessage << "Average mflops/s per call per node : " << mflops/NN << std::endl;
RealD Fullmflops = 1154*volume*DhopCalls/(DhopTotalTime)/2; // 2 for red black counting
std::cout << GridLogMessage << "Average mflops/s per call (full) : " << Fullmflops << std::endl;
std::cout << GridLogMessage << "Average mflops/s per call per rank (full): " << Fullmflops/NP << std::endl;
std::cout << GridLogMessage << "Average mflops/s per call per node (full): " << Fullmflops/NN << std::endl;
std::cout << GridLogMessage << "ImprovedStaggeredFermion5D Stencil" <<std::endl; Stencil.Report();
std::cout << GridLogMessage << "ImprovedStaggeredFermion5D StencilEven"<<std::endl; StencilEven.Report();
std::cout << GridLogMessage << "ImprovedStaggeredFermion5D StencilOdd" <<std::endl; StencilOdd.Report();
}
template<class Impl>
void ImprovedStaggeredFermion5D<Impl>::ZeroCounters(void)
{
DhopCalls = 0;
DhopTotalTime = 0;
DhopCommTime = 0;
DhopComputeTime = 0;
Stencil.ZeroCounters();
StencilEven.ZeroCounters();
StencilOdd.ZeroCounters();
}
/////////////////////////////////////////////////////////////////////////
// Implement the general interface. Here we use SAME mass on all slices

10
lib/qcd/action/fermion/ImprovedStaggeredFermion5D.h
View File

@ -55,6 +55,16 @@ namespace QCD {
FermionField _tmp;
FermionField &tmp(void) { return _tmp; }
////////////////////////////////////////
// Performance monitoring
////////////////////////////////////////
void Report(void);
void ZeroCounters(void);
double DhopTotalTime;
double DhopCalls;
double DhopCommTime;
double DhopComputeTime;
///////////////////////////////////////////////////////////////
// Implement the abstract base
///////////////////////////////////////////////////////////////

48
lib/qcd/action/fermion/WilsonCompressor.h
View File

@ -238,7 +238,33 @@ template<typename HCS,typename HS,typename S> using WilsonCompressor = WilsonCom
template<class vobj,class cobj>
class WilsonStencil : public CartesianStencil<vobj,cobj> {
public:
double timer0;
double timer1;
double timer2;
double timer3;
double timer4;
double timer5;
double timer6;
uint64_t callsi;
void ZeroCountersi(void)
{
timer0=0;
timer1=0;
timer2=0;
timer3=0;
timer4=0;
timer5=0;
timer6=0;
callsi=0;
}
void Reporti(int calls)
{
if ( timer0 ) std::cout << GridLogMessage << " timer0 (HaloGatherOpt) " <<timer0/calls <<std::endl;
if ( timer1 ) std::cout << GridLogMessage << " timer1 (Communicate) " <<timer1/calls <<std::endl;
if ( timer2 ) std::cout << GridLogMessage << " timer2 (CommsMerge ) " <<timer2/calls <<std::endl;
if ( timer3 ) std::cout << GridLogMessage << " timer3 (commsMergeShm) " <<timer3/calls <<std::endl;
if ( timer4 ) std::cout << GridLogMessage << " timer4 " <<timer4 <<std::endl;
}
typedef CartesianCommunicator::CommsRequest_t CommsRequest_t;
std::vector<int> same_node;
@ -252,6 +278,7 @@ public:
: CartesianStencil<vobj,cobj> (grid,npoints,checkerboard,directions,distances) ,
same_node(npoints)
{
ZeroCountersi();
surface_list.resize(0);
};
@ -261,7 +288,6 @@ public:
// Here we know the distance is 1 for WilsonStencil
for(int point=0;point<this->_npoints;point++){
same_node[point] = this->SameNode(point);
// std::cout << " dir " <<point<<" same_node " <<same_node[point]<<std::endl;
}
for(int site = 0 ;site< vol4;site++){
@ -282,17 +308,28 @@ public:
{
std::vector<std::vector<CommsRequest_t> > reqs;
this->HaloExchangeOptGather(source,compress);
this->CommunicateBegin(reqs);
this->CommunicateComplete(reqs);
double t1=usecond();
// Asynchronous MPI calls multidirectional, Isend etc...
// this->CommunicateBegin(reqs);
// this->CommunicateComplete(reqs);
// Non-overlapped directions within a thread. Asynchronous calls except MPI3, threaded up to comm threads ways.
this->Communicate();
double t2=usecond(); timer1 += t2-t1;
this->CommsMerge(compress);
double t3=usecond(); timer2 += t3-t2;
this->CommsMergeSHM(compress);
double t4=usecond(); timer3 += t4-t3;
}
template <class compressor>
void HaloExchangeOptGather(const Lattice<vobj> &source,compressor &compress)
{
this->Prepare();
double t0=usecond();
this->HaloGatherOpt(source,compress);
double t1=usecond();
timer0 += t1-t0;
callsi++;
}
template <class compressor>
@ -304,7 +341,9 @@ public:
typedef typename compressor::SiteHalfSpinor SiteHalfSpinor;
typedef typename compressor::SiteHalfCommSpinor SiteHalfCommSpinor;
this->mpi3synctime_g-=usecond();
this->_grid->StencilBarrier();
this->mpi3synctime_g+=usecond();
assert(source._grid==this->_grid);
this->halogtime-=usecond();
@ -323,7 +362,6 @@ public:
int dag = compress.dag;
int face_idx=0;
if ( dag ) {
// std::cout << " Optimised Dagger compress " <<std::endl;
assert(same_node[Xp]==this->HaloGatherDir(source,XpCompress,Xp,face_idx));
assert(same_node[Yp]==this->HaloGatherDir(source,YpCompress,Yp,face_idx));
assert(same_node[Zp]==this->HaloGatherDir(source,ZpCompress,Zp,face_idx));

107
lib/qcd/action/fermion/WilsonFermion5D.cc
View File

@ -123,22 +123,24 @@ WilsonFermion5D<Impl>::WilsonFermion5D(GaugeField &_Umu,
int vol4;
vol4=FourDimGrid.oSites();
Stencil.BuildSurfaceList(LLs,vol4);
vol4=FourDimRedBlackGrid.oSites();
StencilEven.BuildSurfaceList(LLs,vol4);
StencilOdd.BuildSurfaceList(LLs,vol4);
std::cout << GridLogMessage << " SurfaceLists "<< Stencil.surface_list.size()
<<" " << StencilEven.surface_list.size()<<std::endl;
// std::cout << GridLogMessage << " SurfaceLists "<< Stencil.surface_list.size()
// <<" " << StencilEven.surface_list.size()<<std::endl;
}
template<class Impl>
void WilsonFermion5D<Impl>::Report(void)
{
std::vector<int> latt = GridDefaultLatt();
RealD volume = Ls; for(int mu=0;mu<Nd;mu++) volume=volume*latt[mu];
RealD NP = _FourDimGrid->_Nprocessors;
RealD NN = _FourDimGrid->NodeCount();
RealD NP = _FourDimGrid->_Nprocessors;
RealD NN = _FourDimGrid->NodeCount();
RealD volume = Ls;
std::vector<int> latt = _FourDimGrid->GlobalDimensions();
for(int mu=0;mu<Nd;mu++) volume=volume*latt[mu];
if ( DhopCalls > 0 ) {
std::cout << GridLogMessage << "#### Dhop calls report " << std::endl;
@ -184,6 +186,11 @@ void WilsonFermion5D<Impl>::Report(void)
std::cout << GridLogMessage << "WilsonFermion5D StencilEven"<<std::endl; StencilEven.Report();
std::cout << GridLogMessage << "WilsonFermion5D StencilOdd" <<std::endl; StencilOdd.Report();
}
if ( DhopCalls > 0){
std::cout << GridLogMessage << "WilsonFermion5D Stencil Reporti()" <<std::endl; Stencil.Reporti(DhopCalls);
std::cout << GridLogMessage << "WilsonFermion5D StencilEven Reporti()"<<std::endl; StencilEven.Reporti(DhopCalls);
std::cout << GridLogMessage << "WilsonFermion5D StencilOdd Reporti()" <<std::endl; StencilOdd.Reporti(DhopCalls);
}
}
template<class Impl>
@ -203,6 +210,9 @@ void WilsonFermion5D<Impl>::ZeroCounters(void) {
Stencil.ZeroCounters();
StencilEven.ZeroCounters();
StencilOdd.ZeroCounters();
Stencil.ZeroCountersi();
StencilEven.ZeroCountersi();
StencilOdd.ZeroCountersi();
}
@ -379,7 +389,6 @@ void WilsonFermion5D<Impl>::DhopInternalOverlappedComms(StencilImpl & st, Lebesg
{
#ifdef GRID_OMP
// assert((dag==DaggerNo) ||(dag==DaggerYes));
typedef CartesianCommunicator::CommsRequest_t CommsRequest_t;
Compressor compressor(dag);
@ -388,46 +397,70 @@ void WilsonFermion5D<Impl>::DhopInternalOverlappedComms(StencilImpl & st, Lebesg
DhopFaceTime-=usecond();
st.HaloExchangeOptGather(in,compressor);
DhopFaceTime+=usecond();
std::vector<std::vector<CommsRequest_t> > reqs;
// Rely on async comms; start comms before merge of local data
DhopCommTime-=usecond();
st.CommunicateBegin(reqs);
DhopFaceTime-=usecond();
st.CommsMergeSHM(compressor);
st.CommsMergeSHM(compressor);// Could do this inside parallel region overlapped with comms
DhopFaceTime+=usecond();
// Perhaps use omp task and region
#pragma omp parallel
double ctime=0;
double ptime=0;
//////////////////////////////////////////////////////////////////////////////////////////////////////
// Ugly explicit thread mapping introduced for OPA reasons.
//////////////////////////////////////////////////////////////////////////////////////////////////////
#pragma omp parallel reduction(max:ctime) reduction(max:ptime)
{
int tid = omp_get_thread_num();
int nthreads = omp_get_num_threads();
int me = omp_get_thread_num();
int myoff, mywork;
GridThread::GetWork(len,me-1,mywork,myoff,nthreads-1);
int sF = LLs * myoff;
if ( me == 0 ) {
st.CommunicateComplete(reqs);
DhopCommTime+=usecond();
} else {
// Interior links in stencil
if ( me==1 ) DhopComputeTime-=usecond();
if (dag == DaggerYes) Kernels::DhopSiteDag(st,lo,U,st.CommBuf(),sF,myoff,LLs,mywork,in,out,1,0);
else Kernels::DhopSite(st,lo,U,st.CommBuf(),sF,myoff,LLs,mywork,in,out,1,0);
if ( me==1 ) DhopComputeTime+=usecond();
int ncomms = CartesianCommunicator::nCommThreads;
if (ncomms == -1) ncomms = 1;
assert(nthreads > ncomms);
if (tid >= ncomms) {
double start = usecond();
nthreads -= ncomms;
int ttid = tid - ncomms;
int n = U._grid->oSites();
int chunk = n / nthreads;
int rem = n % nthreads;
int myblock, myn;
if (ttid < rem) {
myblock = ttid * chunk + ttid;
myn = chunk+1;
} else {
myblock = ttid*chunk + rem;
myn = chunk;
}
// do the compute
if (dag == DaggerYes) {
for (int ss = myblock; ss < myblock+myn; ++ss) {
int sU = ss;
int sF = LLs * sU;
Kernels::DhopSiteDag(st,lo,U,st.CommBuf(),sF,sU,LLs,1,in,out,1,0);
}
} else {
for (int ss = myblock; ss < myblock+myn; ++ss) {
int sU = ss;
int sF = LLs * sU;
Kernels::DhopSite(st,lo,U,st.CommBuf(),sF,sU,LLs,1,in,out,1,0);
}
}
ptime = usecond() - start;
}
{
double start = usecond();
st.CommunicateThreaded();
ctime = usecond() - start;
}
}
DhopCommTime += ctime;
DhopComputeTime+=ptime;
// First to enter, last to leave timing
st.CollateThreads();
DhopFaceTime-=usecond();
st.CommsMerge(compressor);
DhopFaceTime+=usecond();
// Load imbalance alert. Should use dynamic schedule OMP for loop
// Perhaps create a list of only those sites with face work, and
// load balance process the list.
DhopComputeTime2-=usecond();
if (dag == DaggerYes) {
int sz=st.surface_list.size();
@ -448,11 +481,9 @@ void WilsonFermion5D<Impl>::DhopInternalOverlappedComms(StencilImpl & st, Lebesg
#else
assert(0);
#endif
}
template<class Impl>
void WilsonFermion5D<Impl>::DhopInternalSerialComms(StencilImpl & st, LebesgueOrder &lo,
DoubledGaugeField & U,

20
lib/qcd/action/gauge/GaugeImplTypes.h
View File

@ -40,12 +40,15 @@ namespace QCD {
typedef typename GImpl::Simd Simd; \
typedef typename GImpl::LinkField GaugeLinkField; \
typedef typename GImpl::Field GaugeField; \
typedef typename GImpl::ComplexField ComplexField;\
typedef typename GImpl::SiteField SiteGaugeField; \
typedef typename GImpl::SiteComplex SiteComplex; \
typedef typename GImpl::SiteLink SiteGaugeLink;
#define INHERIT_FIELD_TYPES(Impl) \
typedef typename Impl::Simd Simd; \
typedef typename Impl::SiteField SiteField; \
#define INHERIT_FIELD_TYPES(Impl) \
typedef typename Impl::Simd Simd; \
typedef typename Impl::ComplexField ComplexField; \
typedef typename Impl::SiteField SiteField; \
typedef typename Impl::Field Field;
// hardcodes the exponential approximation in the template
@ -53,14 +56,17 @@ template <class S, int Nrepresentation = Nc, int Nexp = 12 > class GaugeImplType
public:
typedef S Simd;
template <typename vtype> using iImplGaugeLink = iScalar<iScalar<iMatrix<vtype, Nrepresentation>>>;
template <typename vtype> using iImplGaugeField = iVector<iScalar<iMatrix<vtype, Nrepresentation>>, Nd>;
template <typename vtype> using iImplScalar = iScalar<iScalar<iScalar<vtype> > >;
template <typename vtype> using iImplGaugeLink = iScalar<iScalar<iMatrix<vtype, Nrepresentation> > >;
template <typename vtype> using iImplGaugeField = iVector<iScalar<iMatrix<vtype, Nrepresentation> >, Nd>;
typedef iImplScalar<Simd> SiteComplex;
typedef iImplGaugeLink<Simd> SiteLink;
typedef iImplGaugeField<Simd> SiteField;
typedef Lattice<SiteLink> LinkField;
typedef Lattice<SiteField> Field;
typedef Lattice<SiteComplex> ComplexField;
typedef Lattice<SiteLink> LinkField;
typedef Lattice<SiteField> Field;
// Guido: we can probably separate the types from the HMC functions
// this will create 2 kind of implementations

286
lib/qcd/action/gauge/Photon.h Normal file
View File

@ -0,0 +1,286 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/qcd/action/gauge/Photon.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 QCD_PHOTON_ACTION_H
#define QCD_PHOTON_ACTION_H
namespace Grid{
namespace QCD{
template <class S>
class QedGimpl
{
public:
typedef S Simd;
template <typename vtype>
using iImplGaugeLink = iScalar<iScalar<iScalar<vtype>>>;
template <typename vtype>
using iImplGaugeField = iVector<iScalar<iScalar<vtype>>, Nd>;
typedef iImplGaugeLink<Simd> SiteLink;
typedef iImplGaugeField<Simd> SiteField;
typedef SiteField SiteComplex;
typedef Lattice<SiteLink> LinkField;
typedef Lattice<SiteField> Field;
typedef Field ComplexField;
};
typedef QedGimpl<vComplex> QedGimplR;
template<class Gimpl>
class Photon
{
public:
INHERIT_GIMPL_TYPES(Gimpl);
GRID_SERIALIZABLE_ENUM(Gauge, undef, feynman, 1, coulomb, 2, landau, 3);
GRID_SERIALIZABLE_ENUM(ZmScheme, undef, qedL, 1, qedTL, 2);
public:
Photon(Gauge gauge, ZmScheme zmScheme);
virtual ~Photon(void) = default;
void FreePropagator(const GaugeField &in, GaugeField &out);
void MomentumSpacePropagator(const GaugeField &in, GaugeField &out);
void StochasticWeight(GaugeLinkField &weight);
void StochasticField(GaugeField &out, GridParallelRNG &rng);
void StochasticField(GaugeField &out, GridParallelRNG &rng,
const GaugeLinkField &weight);
private:
void invKHatSquared(GaugeLinkField &out);
void zmSub(GaugeLinkField &out);
private:
Gauge gauge_;
ZmScheme zmScheme_;
};
typedef Photon<QedGimplR> PhotonR;
template<class Gimpl>
Photon<Gimpl>::Photon(Gauge gauge, ZmScheme zmScheme)
: gauge_(gauge), zmScheme_(zmScheme)
{}
template<class Gimpl>
void Photon<Gimpl>::FreePropagator (const GaugeField &in,GaugeField &out)
{
FFT theFFT(in._grid);
GaugeField in_k(in._grid);
GaugeField prop_k(in._grid);
theFFT.FFT_all_dim(in_k,in,FFT::forward);
MomentumSpacePropagator(prop_k,in_k);
theFFT.FFT_all_dim(out,prop_k,FFT::backward);
}
template<class Gimpl>
void Photon<Gimpl>::invKHatSquared(GaugeLinkField &out)
{
GridBase *grid = out._grid;
GaugeLinkField kmu(grid), one(grid);
const unsigned int nd = grid->_ndimension;
std::vector<int> &l = grid->_fdimensions;
std::vector<int> zm(nd,0);
TComplex Tone = Complex(1.0,0.0);
TComplex Tzero= Complex(0.0,0.0);
one = Complex(1.0,0.0);
out = zero;
for(int mu = 0; mu < nd; mu++)
{
Real twoPiL = M_PI*2./l[mu];
LatticeCoordinate(kmu,mu);
kmu = 2.*sin(.5*twoPiL*kmu);
out = out + kmu*kmu;
}
pokeSite(Tone, out, zm);
out = one/out;
pokeSite(Tzero, out, zm);
}
template<class Gimpl>
void Photon<Gimpl>::zmSub(GaugeLinkField &out)
{
GridBase *grid = out._grid;
const unsigned int nd = grid->_ndimension;
switch (zmScheme_)
{
case ZmScheme::qedTL:
{
std::vector<int> zm(nd,0);
TComplex Tzero = Complex(0.0,0.0);
pokeSite(Tzero, out, zm);
break;
}
case ZmScheme::qedL:
{
LatticeInteger spNrm(grid), coor(grid);
GaugeLinkField z(grid);
spNrm = zero;
for(int d = 0; d < grid->_ndimension - 1; d++)
{
LatticeCoordinate(coor,d);
spNrm = spNrm + coor*coor;
}
out = where(spNrm == Integer(0), 0.*out, out);
break;
}
default:
break;
}
}
template<class Gimpl>
void Photon<Gimpl>::MomentumSpacePropagator(const GaugeField &in,
GaugeField &out)
{
GridBase *grid = out._grid;
LatticeComplex k2Inv(grid);
invKHatSquared(k2Inv);
zmSub(k2Inv);
out = in*k2Inv;
}
template<class Gimpl>
void Photon<Gimpl>::StochasticWeight(GaugeLinkField &weight)
{
auto *grid = dynamic_cast<GridCartesian *>(weight._grid);
const unsigned int nd = grid->_ndimension;
std::vector<int> latt_size = grid->_fdimensions;
Integer vol = 1;
for(int d = 0; d < nd; d++)
{
vol = vol * latt_size[d];
}
invKHatSquared(weight);
weight = sqrt(vol*real(weight));
zmSub(weight);
}
template<class Gimpl>
void Photon<Gimpl>::StochasticField(GaugeField &out, GridParallelRNG &rng)
{
auto *grid = dynamic_cast<GridCartesian *>(out._grid);
GaugeLinkField weight(grid);
StochasticWeight(weight);
StochasticField(out, rng, weight);
}
template<class Gimpl>
void Photon<Gimpl>::StochasticField(GaugeField &out, GridParallelRNG &rng,
const GaugeLinkField &weight)
{
auto *grid = dynamic_cast<GridCartesian *>(out._grid);
const unsigned int nd = grid->_ndimension;
GaugeLinkField r(grid);
GaugeField aTilde(grid);
FFT fft(grid);
for(int mu = 0; mu < nd; mu++)
{
gaussian(rng, r);
r = weight*r;
pokeLorentz(aTilde, r, mu);
}
fft.FFT_all_dim(out, aTilde, FFT::backward);
out = real(out);
}
// template<class Gimpl>
// void Photon<Gimpl>::FeynmanGaugeMomentumSpacePropagator_L(GaugeField &out,
// const GaugeField &in)
// {
//
// FeynmanGaugeMomentumSpacePropagator_TL(out,in);
//
// GridBase *grid = out._grid;
// LatticeInteger coor(grid);
// GaugeField zz(grid); zz=zero;
//
// // xyzt
// for(int d = 0; d < grid->_ndimension-1;d++){
// LatticeCoordinate(coor,d);
// out = where(coor==Integer(0),zz,out);
// }
// }
//
// template<class Gimpl>
// void Photon<Gimpl>::FeynmanGaugeMomentumSpacePropagator_TL(GaugeField &out,
// const GaugeField &in)
// {
//
// // what type LatticeComplex
// GridBase *grid = out._grid;
// int nd = grid->_ndimension;
//
// typedef typename GaugeField::vector_type vector_type;
// typedef typename GaugeField::scalar_type ScalComplex;
// typedef Lattice<iSinglet<vector_type> > LatComplex;
//
// std::vector<int> latt_size = grid->_fdimensions;
//
// LatComplex denom(grid); denom= zero;
// LatComplex one(grid); one = ScalComplex(1.0,0.0);
// LatComplex kmu(grid);
//
// ScalComplex ci(0.0,1.0);
// // momphase = n * 2pi / L
// for(int mu=0;mu<Nd;mu++) {
//
// LatticeCoordinate(kmu,mu);
//
// RealD TwoPiL = M_PI * 2.0/ latt_size[mu];
//
// kmu = TwoPiL * kmu ;
//
// denom = denom + 4.0*sin(kmu*0.5)*sin(kmu*0.5); // Wilson term
// }
// std::vector<int> zero_mode(nd,0);
// TComplexD Tone = ComplexD(1.0,0.0);
// TComplexD Tzero= ComplexD(0.0,0.0);
//
// pokeSite(Tone,denom,zero_mode);
//
// denom= one/denom;
//
// pokeSite(Tzero,denom,zero_mode);
//
// out = zero;
// out = in*denom;
// };
}}
#endif

5
lib/qcd/action/scalar/Scalar.h
View File

@ -31,6 +31,7 @@ directory
#include <Grid/qcd/action/scalar/ScalarImpl.h>
#include <Grid/qcd/action/scalar/ScalarAction.h>
#include <Grid/qcd/action/scalar/ScalarInteractionAction.h>
namespace Grid {
namespace QCD {
@ -39,6 +40,10 @@ namespace QCD {
typedef ScalarAction<ScalarImplF> ScalarActionF;
typedef ScalarAction<ScalarImplD> ScalarActionD;
template <int Colours, int Dimensions> using ScalarAdjActionR = ScalarInteractionAction<ScalarNxNAdjImplR<Colours>, Dimensions>;
template <int Colours, int Dimensions> using ScalarAdjActionF = ScalarInteractionAction<ScalarNxNAdjImplF<Colours>, Dimensions>;
template <int Colours, int Dimensions> using ScalarAdjActionD = ScalarInteractionAction<ScalarNxNAdjImplD<Colours>, Dimensions>;
}
}

61
lib/qcd/action/scalar/ScalarAction.h
View File

@ -6,10 +6,10 @@
Copyright (C) 2015
Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk>
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
Author: neo <cossu@post.kek.jp>
Author: paboyle <paboyle@ph.ed.ac.uk>
Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk>
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
Author: neo <cossu@post.kek.jp>
Author: paboyle <paboyle@ph.ed.ac.uk>
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
@ -35,50 +35,49 @@ directory
namespace Grid {
// FIXME drop the QCD namespace everywhere here
template <class Impl>
class ScalarAction : public QCD::Action<typename Impl::Field> {
public:
template <class Impl>
class ScalarAction : public QCD::Action<typename Impl::Field> {
public:
INHERIT_FIELD_TYPES(Impl);
private:
private:
RealD mass_square;
RealD lambda;
public:
ScalarAction(RealD ms, RealD l) : mass_square(ms), lambda(l){};
virtual std::string LogParameters(){
public:
ScalarAction(RealD ms, RealD l) : mass_square(ms), lambda(l) {}
virtual std::string LogParameters() {
std::stringstream sstream;
sstream << GridLogMessage << "[ScalarAction] lambda : " << lambda << std::endl;
sstream << GridLogMessage << "[ScalarAction] mass_square : " << mass_square << std::endl;
return sstream.str();
}
virtual std::string action_name(){return "ScalarAction";}
virtual void refresh(const Field &U,
GridParallelRNG &pRNG){}; // noop as no pseudoferms
virtual std::string action_name() {return "ScalarAction";}
virtual void refresh(const Field &U, GridParallelRNG &pRNG) {} // noop as no pseudoferms
virtual RealD S(const Field &p) {
return (mass_square * 0.5 + QCD::Nd) * ScalarObs<Impl>::sumphisquared(p) +
(lambda / 24.) * ScalarObs<Impl>::sumphifourth(p) +
ScalarObs<Impl>::sumphider(p);
(lambda / 24.) * ScalarObs<Impl>::sumphifourth(p) +
ScalarObs<Impl>::sumphider(p);
};
virtual void deriv(const Field &p,
Field &force) {
Field &force) {
Field tmp(p._grid);
Field p2(p._grid);
ScalarObs<Impl>::phisquared(p2, p);
tmp = -(Cshift(p, 0, -1) + Cshift(p, 0, 1));
for (int mu = 1; mu < QCD::Nd; mu++) tmp -= Cshift(p, mu, -1) + Cshift(p, mu, 1);
force=+(mass_square + 2. * QCD::Nd) * p + (lambda / 6.) * p2 * p + tmp;
};
};
} // Grid
force =+(mass_square + 2. * QCD::Nd) * p + (lambda / 6.) * p2 * p + tmp;
}
};
} // namespace Grid
#endif // SCALAR_ACTION_H

144
lib/qcd/action/scalar/ScalarImpl.h
View File

@ -5,96 +5,158 @@
namespace Grid {
//namespace QCD {
template <class S>
class ScalarImplTypes {
public:
template <class S>
class ScalarImplTypes {
public:
typedef S Simd;
template <typename vtype>
using iImplField = iScalar<iScalar<iScalar<vtype> > >;
typedef iImplField<Simd> SiteField;
typedef SiteField SitePropagator;
typedef SiteField SiteComplex;
typedef Lattice<SiteField> Field;
typedef Field ComplexField;
typedef Field FermionField;
typedef Field PropagatorField;
static inline void generate_momenta(Field& P, GridParallelRNG& pRNG){
gaussian(pRNG, P);
}
static inline Field projectForce(Field& P){return P;}
static inline void update_field(Field& P, Field& U, double ep){
static inline void update_field(Field& P, Field& U, double ep) {
U += P*ep;
}
static inline RealD FieldSquareNorm(Field& U){
static inline RealD FieldSquareNorm(Field& U) {
return (- sum(trace(U*U))/2.0);
}
static inline void HotConfiguration(GridParallelRNG &pRNG, Field &U) {
gaussian(pRNG, U);
}
static inline void TepidConfiguration(GridParallelRNG &pRNG, Field &U) {
gaussian(pRNG, U);
}
static inline void ColdConfiguration(GridParallelRNG &pRNG, Field &U) {
U = 1.0;
}
static void MomentumSpacePropagator(Field &out, RealD m)
{
GridBase *grid = out._grid;
Field kmu(grid), one(grid);
const unsigned int nd = grid->_ndimension;
std::vector<int> &l = grid->_fdimensions;
one = Complex(1.0,0.0);
out = m*m;
for(int mu = 0; mu < nd; mu++)
{
Real twoPiL = M_PI*2./l[mu];
LatticeCoordinate(kmu,mu);
kmu = 2.*sin(.5*twoPiL*kmu);
out = out + kmu*kmu;
}
out = one/out;
}
static void FreePropagator(const Field &in, Field &out,
const Field &momKernel)
{
FFT fft((GridCartesian *)in._grid);
Field inFT(in._grid);
fft.FFT_all_dim(inFT, in, FFT::forward);
inFT = inFT*momKernel;
fft.FFT_all_dim(out, inFT, FFT::backward);
}
static void FreePropagator(const Field &in, Field &out, RealD m)
{
Field momKernel(in._grid);
MomentumSpacePropagator(momKernel, m);
FreePropagator(in, out, momKernel);
}
};
template <class S, unsigned int N>
class ScalarMatrixImplTypes {
class ScalarAdjMatrixImplTypes {
public:
typedef S Simd;
typedef QCD::SU<N> Group;
template <typename vtype>
using iImplField = iScalar<iScalar<iMatrix<vtype, N> > >;
using iImplField = iScalar<iScalar<iMatrix<vtype, N>>>;
template <typename vtype>
using iImplComplex = iScalar<iScalar<iScalar<vtype>>>;
typedef iImplField<Simd> SiteField;
typedef SiteField SitePropagator;
typedef iImplComplex<Simd> SiteComplex;
typedef iImplField<Simd> SiteField;
typedef Lattice<SiteField> Field;
static inline void generate_momenta(Field& P, GridParallelRNG& pRNG){
gaussian(pRNG, P);
typedef Lattice<SiteField> Field;
typedef Lattice<SiteComplex> ComplexField;
typedef Field FermionField;
typedef Field PropagatorField;
static inline void generate_momenta(Field& P, GridParallelRNG& pRNG) {
Group::GaussianFundamentalLieAlgebraMatrix(pRNG, P);
}
static inline Field projectForce(Field& P){return P;}
static inline void update_field(Field& P, Field& U, double ep){
static inline Field projectForce(Field& P) {return P;}
static inline void update_field(Field& P, Field& U, double ep) {
U += P*ep;
}
static inline RealD FieldSquareNorm(Field& U){
return (TensorRemove(- sum(trace(U*U))*0.5).real());
static inline RealD FieldSquareNorm(Field& U) {
return (TensorRemove(sum(trace(U*U))).real());
}
static inline void HotConfiguration(GridParallelRNG &pRNG, Field &U) {
gaussian(pRNG, U);
Group::GaussianFundamentalLieAlgebraMatrix(pRNG, U);
}
static inline void TepidConfiguration(GridParallelRNG &pRNG, Field &U) {
gaussian(pRNG, U);
Group::GaussianFundamentalLieAlgebraMatrix(pRNG, U, 0.01);
}
static inline void ColdConfiguration(GridParallelRNG &pRNG, Field &U) {
U = 1.0;
U = zero;
}
};
typedef ScalarImplTypes<vReal> ScalarImplR;
typedef ScalarImplTypes<vRealF> ScalarImplF;
typedef ScalarImplTypes<vRealD> ScalarImplD;
typedef ScalarImplTypes<vComplex> ScalarImplCR;
typedef ScalarImplTypes<vComplexF> ScalarImplCF;
typedef ScalarImplTypes<vComplexD> ScalarImplCD;
// Hardcoding here the size of the matrices
typedef ScalarAdjMatrixImplTypes<vComplex, QCD::Nc> ScalarAdjImplR;
typedef ScalarAdjMatrixImplTypes<vComplexF, QCD::Nc> ScalarAdjImplF;
typedef ScalarAdjMatrixImplTypes<vComplexD, QCD::Nc> ScalarAdjImplD;
template <int Colours > using ScalarNxNAdjImplR = ScalarAdjMatrixImplTypes<vComplex, Colours >;
template <int Colours > using ScalarNxNAdjImplF = ScalarAdjMatrixImplTypes<vComplexF, Colours >;
template <int Colours > using ScalarNxNAdjImplD = ScalarAdjMatrixImplTypes<vComplexD, Colours >;
//}
}
//}
}
#endif

134
lib/qcd/action/scalar/ScalarInteractionAction.h
View File

@ -6,10 +6,7 @@
Copyright (C) 2015
Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk>
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
Author: neo <cossu@post.kek.jp>
Author: paboyle <paboyle@ph.ed.ac.uk>
Author: Guido Cossu <guido,cossu@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
@ -30,55 +27,122 @@ directory
*************************************************************************************/
/* END LEGAL */
#ifndef SCALAR_ACTION_H
#define SCALAR_ACTION_H
#ifndef SCALAR_INT_ACTION_H
#define SCALAR_INT_ACTION_H
// Note: this action can completely absorb the ScalarAction for real float fields
// use the scalarObjs to generalise the structure
namespace Grid {
// FIXME drop the QCD namespace everywhere here
template <class Impl>
template <class Impl, int Ndim >
class ScalarInteractionAction : public QCD::Action<typename Impl::Field> {
public:
INHERIT_FIELD_TYPES(Impl);
private:
RealD mass_square;
RealD lambda;
public:
ScalarAction(RealD ms, RealD l) : mass_square(ms), lambda(l){};
virtual std::string LogParameters(){
typedef typename Field::vector_object vobj;
typedef CartesianStencil<vobj,vobj> Stencil;
SimpleCompressor<vobj> compressor;
int npoint = 2*Ndim;
std::vector<int> directions;// = {0,1,2,3,0,1,2,3}; // forcing 4 dimensions
std::vector<int> displacements;// = {1,1,1,1, -1,-1,-1,-1};
public:
ScalarInteractionAction(RealD ms, RealD l) : mass_square(ms), lambda(l), displacements(2*Ndim,0), directions(2*Ndim,0){
for (int mu = 0 ; mu < Ndim; mu++){
directions[mu] = mu; directions[mu+Ndim] = mu;
displacements[mu] = 1; displacements[mu+Ndim] = -1;
}
}
virtual std::string LogParameters() {
std::stringstream sstream;
sstream << GridLogMessage << "[ScalarAction] lambda : " << lambda << std::endl;
sstream << GridLogMessage << "[ScalarAction] mass_square : " << mass_square << std::endl;
return sstream.str();
}
virtual std::string action_name(){return "ScalarAction";}
virtual void refresh(const Field &U,
GridParallelRNG &pRNG){}; // noop as no pseudoferms
virtual std::string action_name() {return "ScalarAction";}
virtual void refresh(const Field &U, GridParallelRNG &pRNG) {}
virtual RealD S(const Field &p) {
return (mass_square * 0.5 + QCD::Nd) * ScalarObs<Impl>::sumphisquared(p) +
(lambda / 24.) * ScalarObs<Impl>::sumphifourth(p) +
ScalarObs<Impl>::sumphider(p);
assert(p._grid->Nd() == Ndim);
static Stencil phiStencil(p._grid, npoint, 0, directions, displacements);
phiStencil.HaloExchange(p, compressor);
Field action(p._grid), pshift(p._grid), phisquared(p._grid);
phisquared = p*p;
action = (2.0*Ndim + mass_square)*phisquared - lambda/24.*phisquared*phisquared;
for (int mu = 0; mu < Ndim; mu++) {
// pshift = Cshift(p, mu, +1); // not efficient, implement with stencils
parallel_for (int i = 0; i < p._grid->oSites(); i++) {
int permute_type;
StencilEntry *SE;
vobj temp2;
const vobj *temp, *t_p;
SE = phiStencil.GetEntry(permute_type, mu, i);
t_p = &p._odata[i];
if ( SE->_is_local ) {
temp = &p._odata[SE->_offset];
if ( SE->_permute ) {
permute(temp2, *temp, permute_type);
action._odata[i] -= temp2*(*t_p) + (*t_p)*temp2;
} else {
action._odata[i] -= (*temp)*(*t_p) + (*t_p)*(*temp);
}
} else {
action._odata[i] -= phiStencil.CommBuf()[SE->_offset]*(*t_p) + (*t_p)*phiStencil.CommBuf()[SE->_offset];
}
}
// action -= pshift*p + p*pshift;
}
// NB the trace in the algebra is normalised to 1/2
// minus sign coming from the antihermitian fields
return -(TensorRemove(sum(trace(action)))).real();
};
virtual void deriv(const Field &p,
Field &force) {
Field tmp(p._grid);
Field p2(p._grid);
ScalarObs<Impl>::phisquared(p2, p);
tmp = -(Cshift(p, 0, -1) + Cshift(p, 0, 1));
for (int mu = 1; mu < QCD::Nd; mu++) tmp -= Cshift(p, mu, -1) + Cshift(p, mu, 1);
virtual void deriv(const Field &p, Field &force) {
assert(p._grid->Nd() == Ndim);
force = (2.0*Ndim + mass_square)*p - lambda/12.*p*p*p;
// move this outside
static Stencil phiStencil(p._grid, npoint, 0, directions, displacements);
phiStencil.HaloExchange(p, compressor);
force=+(mass_square + 2. * QCD::Nd) * p + (lambda / 6.) * p2 * p + tmp;
};
//for (int mu = 0; mu < QCD::Nd; mu++) force -= Cshift(p, mu, -1) + Cshift(p, mu, 1);
for (int point = 0; point < npoint; point++) {
parallel_for (int i = 0; i < p._grid->oSites(); i++) {
const vobj *temp;
vobj temp2;
int permute_type;
StencilEntry *SE;
SE = phiStencil.GetEntry(permute_type, point, i);
if ( SE->_is_local ) {
temp = &p._odata[SE->_offset];
if ( SE->_permute ) {
permute(temp2, *temp, permute_type);
force._odata[i] -= temp2;
} else {
force._odata[i] -= *temp;
}
} else {
force._odata[i] -= phiStencil.CommBuf()[SE->_offset];
}
}
}
}
};
} // Grid
} // namespace Grid
#endif // SCALAR_ACTION_H
#endif // SCALAR_INT_ACTION_H

6
lib/qcd/hmc/GenericHMCrunner.h
View File

@ -207,6 +207,12 @@ using GenericHMCRunnerTemplate = HMCWrapperTemplate<Implementation, Integrator,
typedef HMCWrapperTemplate<ScalarImplR, MinimumNorm2, ScalarFields>
ScalarGenericHMCRunner;
typedef HMCWrapperTemplate<ScalarAdjImplR, MinimumNorm2, ScalarMatrixFields>
ScalarAdjGenericHMCRunner;
template <int Colours>
using ScalarNxNAdjGenericHMCRunner = HMCWrapperTemplate < ScalarNxNAdjImplR<Colours>, MinimumNorm2, ScalarNxNMatrixFields<Colours> >;
} // namespace QCD
} // namespace Grid

2
lib/qcd/hmc/HMC.h
View File

@ -76,7 +76,7 @@ struct HMCparameters: Serializable {
template < class ReaderClass >
void initialize(Reader<ReaderClass> &TheReader){
std::cout << "Reading HMC\n";
std::cout << GridLogMessage << "Reading HMC\n";
read(TheReader, "HMC", *this);
}

17
lib/qcd/hmc/HMCResourceManager.h
View File

@ -165,7 +165,7 @@ class HMCResourceManager {
// Grids
//////////////////////////////////////////////////////////////
void AddGrid(std::string s, GridModule& M) {
void AddGrid(const std::string s, GridModule& M) {
// Check for name clashes
auto search = Grids.find(s);
if (search != Grids.end()) {
@ -174,14 +174,24 @@ class HMCResourceManager {
exit(1);
}
Grids[s] = std::move(M);
std::cout << GridLogMessage << "::::::::::::::::::::::::::::::::::::::::" <<std::endl;
std::cout << GridLogMessage << "HMCResourceManager:" << std::endl;
std::cout << GridLogMessage << "Created grid set with name '" << s << "' and decomposition for the full cartesian " << std::endl;
Grids[s].show_full_decomposition();
std::cout << GridLogMessage << "::::::::::::::::::::::::::::::::::::::::" <<std::endl;
}
// Add a named grid set, 4d shortcut
void AddFourDimGrid(std::string s) {
void AddFourDimGrid(const std::string s) {
GridFourDimModule<vComplex> Mod;
AddGrid(s, Mod);
}
// Add a named grid set, 4d shortcut + tweak simd lanes
void AddFourDimGrid(const std::string s, const std::vector<int> simd_decomposition) {
GridFourDimModule<vComplex> Mod(simd_decomposition);
AddGrid(s, Mod);
}
GridCartesian* GetCartesian(std::string s = "") {
@ -253,6 +263,7 @@ class HMCResourceManager {
template<class T, class... Types>
void AddObservable(Types&&... Args){
ObservablesList.push_back(std::unique_ptr<T>(new T(std::forward<Types>(Args)...)));
ObservablesList.back()->print_parameters();
}
std::vector<HmcObservable<typename ImplementationPolicy::Field>* > GetObservables(){
@ -297,4 +308,4 @@ private:
}
}
#endif // HMC_RESOURCE_MANAGER_H
#endif // HMC_RESOURCE_MANAGER_H

92
lib/qcd/hmc/HMC_GridModules.h
View File

@ -33,28 +33,29 @@ directory
namespace Grid {
// Resources
// Modules for grids
// Modules for grids
// Introduce another namespace HMCModules?
class GridModuleParameters: Serializable{
class GridModuleParameters: Serializable{
public:
GRID_SERIALIZABLE_CLASS_MEMBERS(GridModuleParameters,
std::string, lattice,
std::string, mpi);
std::vector<int> getLattice(){return strToVec<int>(lattice);}
std::vector<int> getMpi() {return strToVec<int>(mpi);}
std::vector<int> getLattice() const {return strToVec<int>(lattice);}
std::vector<int> getMpi() const {return strToVec<int>(mpi);}
void check(){
if (getLattice().size() != getMpi().size()) {
std::cout << GridLogError
void check() const {
if (getLattice().size() != getMpi().size() ) {
std::cout << GridLogError
<< "Error in GridModuleParameters: lattice and mpi dimensions "
"do not match"
<< std::endl;
exit(1);
}
}
}
template <class ReaderClass>
GridModuleParameters(Reader<ReaderClass>& Reader, std::string n = "LatticeGrid"):name(n) {
@ -75,51 +76,94 @@ private:
// Lower level class
class GridModule {
public:
GridCartesian* get_full() {
GridCartesian* get_full() {
std::cout << GridLogDebug << "Getting cartesian in module"<< std::endl;
return grid_.get(); }
GridRedBlackCartesian* get_rb() {
GridRedBlackCartesian* get_rb() {
std::cout << GridLogDebug << "Getting rb-cartesian in module"<< std::endl;
return rbgrid_.get(); }
void set_full(GridCartesian* grid) { grid_.reset(grid); }
void set_rb(GridRedBlackCartesian* rbgrid) { rbgrid_.reset(rbgrid); }
void show_full_decomposition(){ grid_->show_decomposition(); }
void show_rb_decomposition(){ rbgrid_->show_decomposition(); }
protected:
std::unique_ptr<GridCartesian> grid_;
std::unique_ptr<GridRedBlackCartesian> rbgrid_;
};
////////////////////////////////////
// Classes for the user
////////////////////////////////////
// Note: the space time grid should be out of the QCD namespace
template< class vector_type>
class GridFourDimModule : public GridModule {
public:
GridFourDimModule() {
template <class vector_type>
class GridFourDimModule : public GridModule
{
public:
GridFourDimModule()
{
using namespace QCD;
set_full(SpaceTimeGrid::makeFourDimGrid(
GridDefaultLatt(), GridDefaultSimd(4, vector_type::Nsimd()),
GridDefaultLatt(),
GridDefaultSimd(4, vector_type::Nsimd()),
GridDefaultMpi()));
set_rb(SpaceTimeGrid::makeFourDimRedBlackGrid(grid_.get()));
}
GridFourDimModule(GridModuleParameters Params) {
GridFourDimModule(const std::vector<int> tweak_simd)
{
using namespace QCD;
if (tweak_simd.size() != 4)
{
std::cout << GridLogError
<< "Error in GridFourDimModule: SIMD size different from 4"
<< std::endl;
exit(1);
}
// Checks that the product agrees with the expectation
int simd_sum = 1;
for (auto &n : tweak_simd)
simd_sum *= n;
std::cout << GridLogDebug << "TweakSIMD: " << tweak_simd << " Sum: " << simd_sum << std::endl;
if (simd_sum == vector_type::Nsimd())
{
set_full(SpaceTimeGrid::makeFourDimGrid(
GridDefaultLatt(),
tweak_simd,
GridDefaultMpi()));
set_rb(SpaceTimeGrid::makeFourDimRedBlackGrid(grid_.get()));
}
else
{
std::cout << GridLogError
<< "Error in GridFourDimModule: SIMD lanes must sum to "
<< vector_type::Nsimd()
<< std::endl;
}
}
GridFourDimModule(const GridModuleParameters Params)
{
using namespace QCD;
Params.check();
std::vector<int> lattice_v = Params.getLattice();
std::vector<int> mpi_v = Params.getMpi();
if (lattice_v.size() == 4) {
if (lattice_v.size() == 4)
{
set_full(SpaceTimeGrid::makeFourDimGrid(
lattice_v, GridDefaultSimd(4, vector_type::Nsimd()),
lattice_v,
GridDefaultSimd(4, vector_type::Nsimd()),
mpi_v));
set_rb(SpaceTimeGrid::makeFourDimRedBlackGrid(grid_.get()));
} else {
std::cout << GridLogError
<< "Error in GridFourDimModule: lattice dimension different from 4"
<< std::endl;
}
else
{
std::cout << GridLogError
<< "Error in GridFourDimModule: lattice dimension different from 4"
<< std::endl;
exit(1);
}
}

42
lib/qcd/hmc/checkpointers/BinaryCheckpointer.h
View File

@ -62,36 +62,50 @@ class BinaryHmcCheckpointer : public BaseHmcCheckpointer<Impl> {
fout.close();
}
void TrajectoryComplete(int traj, Field &U, GridSerialRNG &sRNG,
GridParallelRNG &pRNG) {
void TrajectoryComplete(int traj, Field &U, GridSerialRNG &sRNG, GridParallelRNG &pRNG) {
if ((traj % Params.saveInterval) == 0) {
std::string config, rng;
this->build_filenames(traj, Params, config, rng);
BinaryIO::BinarySimpleUnmunger<sobj_double, sobj> munge;
uint32_t nersc_csum;
uint32_t scidac_csuma;
uint32_t scidac_csumb;
BinarySimpleUnmunger<sobj_double, sobj> munge;
truncate(rng);
BinaryIO::writeRNGSerial(sRNG, pRNG, rng, 0);
BinaryIO::writeRNG(sRNG, pRNG, rng, 0,nersc_csum,scidac_csuma,scidac_csumb);
truncate(config);
uint32_t csum = BinaryIO::writeObjectParallel<vobj, sobj_double>(
U, config, munge, 0, Params.format);
BinaryIO::writeLatticeObject<vobj, sobj_double>(U, config, munge, 0, Params.format,
nersc_csum,scidac_csuma,scidac_csumb);
std::cout << GridLogMessage << "Written Binary Configuration " << config
<< " checksum " << std::hex << csum << std::dec << std::endl;
<< " checksum " << std::hex
<< nersc_csum <<"/"
<< scidac_csuma <<"/"
<< scidac_csumb
<< std::dec << std::endl;
}
};
void CheckpointRestore(int traj, Field &U, GridSerialRNG &sRNG,
GridParallelRNG &pRNG) {
void CheckpointRestore(int traj, Field &U, GridSerialRNG &sRNG, GridParallelRNG &pRNG) {
std::string config, rng;
this->build_filenames(traj, Params, config, rng);
BinaryIO::BinarySimpleMunger<sobj_double, sobj> munge;
BinaryIO::readRNGSerial(sRNG, pRNG, rng, 0);
uint32_t csum = BinaryIO::readObjectParallel<vobj, sobj_double>(
U, config, munge, 0, Params.format);
BinarySimpleMunger<sobj_double, sobj> munge;
uint32_t nersc_csum;
uint32_t scidac_csuma;
uint32_t scidac_csumb;
BinaryIO::readRNG(sRNG, pRNG, rng, 0,nersc_csum,scidac_csuma,scidac_csumb);
BinaryIO::readLatticeObject<vobj, sobj_double>(U, config, munge, 0, Params.format,
nersc_csum,scidac_csuma,scidac_csumb);
std::cout << GridLogMessage << "Read Binary Configuration " << config
<< " checksum " << std::hex << csum << std::dec << std::endl;
<< " checksums " << std::hex << nersc_csum<<"/"<<scidac_csuma<<"/"<<scidac_csumb
<< std::dec << std::endl;
};
};
}

38
lib/qcd/hmc/checkpointers/ILDGCheckpointer.h
View File

@ -54,9 +54,9 @@ class ILDGHmcCheckpointer : public BaseHmcCheckpointer<Implementation> {
// check here that the format is valid
int ieee32big = (Params.format == std::string("IEEE32BIG"));
int ieee32 = (Params.format == std::string("IEEE32"));
int ieee32 = (Params.format == std::string("IEEE32"));
int ieee64big = (Params.format == std::string("IEEE64BIG"));
int ieee64 = (Params.format == std::string("IEEE64"));
int ieee64 = (Params.format == std::string("IEEE64"));
if (!(ieee64big || ieee32 || ieee32big || ieee64)) {
std::cout << GridLogError << "Unrecognized file format " << Params.format
@ -74,13 +74,20 @@ class ILDGHmcCheckpointer : public BaseHmcCheckpointer<Implementation> {
if ((traj % Params.saveInterval) == 0) {
std::string config, rng;
this->build_filenames(traj, Params, config, rng);
ILDGIO IO(config, ILDGwrite);
BinaryIO::writeRNGSerial(sRNG, pRNG, rng, 0);
uint32_t csum = IO.writeConfiguration(U, Params.format);
uint32_t nersc_csum,scidac_csuma,scidac_csumb;
BinaryIO::writeRNG(sRNG, pRNG, rng, 0,nersc_csum,scidac_csuma,scidac_csumb);
IldgWriter _IldgWriter;
_IldgWriter.open(config);
_IldgWriter.writeConfiguration(U, traj, config, config);
_IldgWriter.close();
std::cout << GridLogMessage << "Written ILDG Configuration on " << config
<< " checksum " << std::hex << csum << std::dec << std::endl;
<< " checksum " << std::hex
<< nersc_csum<<"/"
<< scidac_csuma<<"/"
<< scidac_csumb
<< std::dec << std::endl;
}
};
@ -89,12 +96,21 @@ class ILDGHmcCheckpointer : public BaseHmcCheckpointer<Implementation> {
std::string config, rng;
this->build_filenames(traj, Params, config, rng);
ILDGIO IO(config, ILDGread);
BinaryIO::readRNGSerial(sRNG, pRNG, rng, 0);
uint32_t csum = IO.readConfiguration(U); // format from the header
uint32_t nersc_csum,scidac_csuma,scidac_csumb;
BinaryIO::readRNG(sRNG, pRNG, rng, 0,nersc_csum,scidac_csuma,scidac_csumb);
FieldMetaData header;
IldgReader _IldgReader;
_IldgReader.open(config);
_IldgReader.readConfiguration(U,header); // format from the header
_IldgReader.close();
std::cout << GridLogMessage << "Read ILDG Configuration from " << config
<< " checksum " << std::hex << csum << std::dec << std::endl;
<< " checksum " << std::hex
<< nersc_csum<<"/"
<< scidac_csuma<<"/"
<< scidac_csumb
<< std::dec << std::endl;
};
};
}

2
lib/qcd/hmc/checkpointers/NerscCheckpointer.h
View File

@ -70,7 +70,7 @@ class NerscHmcCheckpointer : public BaseHmcCheckpointer<Gimpl> {
std::string config, rng;
this->build_filenames(traj, Params, config, rng);
NerscField header;
FieldMetaData header;
NerscIO::readRNGState(sRNG, pRNG, header, rng);
NerscIO::readConfiguration(U, header, config);
};

15
lib/qcd/modules/ObservableModules.h
View File

@ -84,8 +84,6 @@ class PlaquetteMod: public ObservableModule<PlaquetteLogger<Impl>, NoParameters>
typedef ObservableModule<PlaquetteLogger<Impl>, NoParameters> ObsBase;
using ObsBase::ObsBase; // for constructors
// acquire resource
virtual void initialize(){
this->ObservablePtr.reset(new PlaquetteLogger<Impl>());
@ -94,23 +92,22 @@ class PlaquetteMod: public ObservableModule<PlaquetteLogger<Impl>, NoParameters>
PlaquetteMod(): ObsBase(NoParameters()){}
};
template < class Impl >
class TopologicalChargeMod: public ObservableModule<TopologicalCharge<Impl>, NoParameters>{
typedef ObservableModule<TopologicalCharge<Impl>, NoParameters> ObsBase;
class TopologicalChargeMod: public ObservableModule<TopologicalCharge<Impl>, TopologyObsParameters>{
typedef ObservableModule<TopologicalCharge<Impl>, TopologyObsParameters> ObsBase;
using ObsBase::ObsBase; // for constructors
// acquire resource
virtual void initialize(){
this->ObservablePtr.reset(new TopologicalCharge<Impl>());
this->ObservablePtr.reset(new TopologicalCharge<Impl>(this->Par_));
}
public:
TopologicalChargeMod(): ObsBase(NoParameters()){}
TopologicalChargeMod(TopologyObsParameters Par): ObsBase(Par){}
TopologicalChargeMod(): ObsBase(){}
};
}// QCD temporarily here

69
lib/qcd/observables/topological_charge.h
View File

@ -33,9 +33,45 @@ directory
namespace Grid {
namespace QCD {
struct TopologySmearingParameters : Serializable {
GRID_SERIALIZABLE_CLASS_MEMBERS(TopologySmearingParameters,
int, steps,
float, step_size,
int, meas_interval,
float, maxTau);
TopologySmearingParameters(int s = 0, float ss = 0.0f, int mi = 0, float mT = 0.0f):
steps(s), step_size(ss), meas_interval(mi), maxTau(mT){}
template < class ReaderClass >
TopologySmearingParameters(Reader<ReaderClass>& Reader){
read(Reader, "Smearing", *this);
}
};
struct TopologyObsParameters : Serializable {
GRID_SERIALIZABLE_CLASS_MEMBERS(TopologyObsParameters,
int, interval,
bool, do_smearing,
TopologySmearingParameters, Smearing);
TopologyObsParameters(int interval = 1, bool smearing = false):
interval(interval), Smearing(smearing){}
template <class ReaderClass >
TopologyObsParameters(Reader<ReaderClass>& Reader){
read(Reader, "TopologyMeasurement", *this);
}
};
// this is only defined for a gauge theory
template <class Impl>
class TopologicalCharge : public HmcObservable<typename Impl::Field> {
TopologyObsParameters Pars;
public:
// here forces the Impl to be of gauge fields
// if not the compiler will complain
@ -44,20 +80,39 @@ class TopologicalCharge : public HmcObservable<typename Impl::Field> {
// necessary for HmcObservable compatibility
typedef typename Impl::Field Field;
TopologicalCharge(int interval = 1, bool do_smearing = false):
Pars(interval, do_smearing){}
TopologicalCharge(TopologyObsParameters P):Pars(P){
std::cout << GridLogDebug << "Creating TopologicalCharge " << std::endl;
}
void TrajectoryComplete(int traj,
Field &U,
GridSerialRNG &sRNG,
GridParallelRNG &pRNG) {
Real q = WilsonLoops<Impl>::TopologicalCharge(U);
if (traj%Pars.interval == 0){
// Smearing
Field Usmear = U;
int def_prec = std::cout.precision();
if (Pars.do_smearing){
// using wilson flow by default here
WilsonFlow<PeriodicGimplR> WF(Pars.Smearing.steps, Pars.Smearing.step_size, Pars.Smearing.meas_interval);
WF.smear_adaptive(Usmear, U, Pars.Smearing.maxTau);
Real T0 = WF.energyDensityPlaquette(Usmear);
std::cout << GridLogMessage << std::setprecision(std::numeric_limits<Real>::digits10 + 1)
<< "T0 : [ " << traj << " ] "<< T0 << std::endl;
}
int def_prec = std::cout.precision();
Real q = WilsonLoops<Impl>::TopologicalCharge(Usmear);
std::cout << GridLogMessage
<< std::setprecision(std::numeric_limits<Real>::digits10 + 1)
<< "Topological Charge: [ " << traj << " ] "<< q << std::endl;
std::cout << GridLogMessage
<< std::setprecision(std::numeric_limits<Real>::digits10 + 1)
<< "Topological Charge: [ " << traj << " ] "<< q << std::endl;
std::cout.precision(def_prec);
std::cout.precision(def_prec);
}
}
};

5
lib/qcd/representations/hmc_types.h
View File

@ -62,7 +62,10 @@ class Representations {
typedef Representations<FundamentalRepresentation> NoHirep;
typedef Representations<EmptyRep<typename ScalarImplR::Field> > ScalarFields;
//typedef Representations<EmptyRep<typename ScalarMatrixImplR::Field> > ScalarMatrixFields;
typedef Representations<EmptyRep<typename ScalarAdjImplR::Field> > ScalarMatrixFields;
template < int Colours>
using ScalarNxNMatrixFields = Representations<EmptyRep<typename ScalarNxNAdjImplR<Colours>::Field> >;
// Helper classes to access the elements
// Strips the first N parameters from the tuple

9
lib/qcd/smearing/WilsonFlow.h
View File

@ -108,7 +108,7 @@ void WilsonFlow<Gimpl>::evolve_step_adaptive(typename Gimpl::GaugeField &U, Real
if (maxTau - taus < epsilon){
epsilon = maxTau-taus;
}
std::cout << GridLogMessage << "Integration epsilon : " << epsilon << std::endl;
//std::cout << GridLogMessage << "Integration epsilon : " << epsilon << std::endl;
GaugeField Z(U._grid);
GaugeField Zprime(U._grid);
GaugeField tmp(U._grid), Uprime(U._grid);
@ -138,10 +138,10 @@ void WilsonFlow<Gimpl>::evolve_step_adaptive(typename Gimpl::GaugeField &U, Real
// adjust integration step
taus += epsilon;
std::cout << GridLogMessage << "Adjusting integration step with distance: " << diff << std::endl;
//std::cout << GridLogMessage << "Adjusting integration step with distance: " << diff << std::endl;
epsilon = epsilon*0.95*std::pow(1e-4/diff,1./3.);
std::cout << GridLogMessage << "New epsilon : " << epsilon << std::endl;
//std::cout << GridLogMessage << "New epsilon : " << epsilon << std::endl;
}
@ -166,7 +166,6 @@ void WilsonFlow<Gimpl>::smear(GaugeField& out, const GaugeField& in) const {
out = in;
for (unsigned int step = 1; step <= Nstep; step++) {
auto start = std::chrono::high_resolution_clock::now();
std::cout << GridLogMessage << "Evolution time :"<< tau(step) << std::endl;
evolve_step(out);
auto end = std::chrono::high_resolution_clock::now();
std::chrono::duration<double> diff = end - start;
@ -191,7 +190,7 @@ void WilsonFlow<Gimpl>::smear_adaptive(GaugeField& out, const GaugeField& in, Re
unsigned int step = 0;
do{
step++;
std::cout << GridLogMessage << "Evolution time :"<< taus << std::endl;
//std::cout << GridLogMessage << "Evolution time :"<< taus << std::endl;
evolve_step_adaptive(out, maxTau);
std::cout << GridLogMessage << "[WilsonFlow] Energy density (plaq) : "
<< step << " "

193
lib/qcd/utils/GaugeFix.h Normal file
View File

@ -0,0 +1,193 @@
/*************************************************************************************
grid` physics library, www.github.com/paboyle/Grid
Copyright (C) 2015
Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk>
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution directory
*************************************************************************************/
/* END LEGAL */
//#include <Grid/Grid.h>
#ifndef GRID_QCD_GAUGE_FIX_H
#define GRID_QCD_GAUGE_FIX_H
namespace Grid {
namespace QCD {
template <class Gimpl>
class FourierAcceleratedGaugeFixer : public Gimpl {
public:
INHERIT_GIMPL_TYPES(Gimpl);
typedef typename Gimpl::GaugeLinkField GaugeMat;
typedef typename Gimpl::GaugeField GaugeLorentz;
static void GaugeLinkToLieAlgebraField(const std::vector<GaugeMat> &U,std::vector<GaugeMat> &A) {
for(int mu=0;mu<Nd;mu++){
Complex cmi(0.0,-1.0);
A[mu] = Ta(U[mu]) * cmi;
}
}
static void DmuAmu(const std::vector<GaugeMat> &A,GaugeMat &dmuAmu) {
dmuAmu=zero;
for(int mu=0;mu<Nd;mu++){
dmuAmu = dmuAmu + A[mu] - Cshift(A[mu],mu,-1);
}
}
static void SteepestDescentGaugeFix(GaugeLorentz &Umu,Real & alpha,int maxiter,Real Omega_tol, Real Phi_tol,bool Fourier=false) {
GridBase *grid = Umu._grid;
Real org_plaq =WilsonLoops<Gimpl>::avgPlaquette(Umu);
Real org_link_trace=WilsonLoops<Gimpl>::linkTrace(Umu);
Real old_trace = org_link_trace;
Real trG;
std::vector<GaugeMat> U(Nd,grid);
GaugeMat dmuAmu(grid);
for(int i=0;i<maxiter;i++){
for(int mu=0;mu<Nd;mu++) U[mu]= PeekIndex<LorentzIndex>(Umu,mu);
if ( Fourier==false ) {
trG = SteepestDescentStep(U,alpha,dmuAmu);
} else {
trG = FourierAccelSteepestDescentStep(U,alpha,dmuAmu);
}
for(int mu=0;mu<Nd;mu++) PokeIndex<LorentzIndex>(Umu,U[mu],mu);
// Monitor progress and convergence test
// infrequently to minimise cost overhead
if ( i %20 == 0 ) {
Real plaq =WilsonLoops<Gimpl>::avgPlaquette(Umu);
Real link_trace=WilsonLoops<Gimpl>::linkTrace(Umu);
if (Fourier)
std::cout << GridLogMessage << "Fourier Iteration "<<i<< " plaq= "<<plaq<< " dmuAmu " << norm2(dmuAmu)<< std::endl;
else
std::cout << GridLogMessage << " Iteration "<<i<< " plaq= "<<plaq<< " dmuAmu " << norm2(dmuAmu)<< std::endl;
Real Phi = 1.0 - old_trace / link_trace ;
Real Omega= 1.0 - trG;
std::cout << GridLogMessage << " Iteration "<<i<< " Phi= "<<Phi<< " Omega= " << Omega<< " trG " << trG <<std::endl;
if ( (Omega < Omega_tol) && ( ::fabs(Phi) < Phi_tol) ) {
std::cout << GridLogMessage << "Converged ! "<<std::endl;
return;
}
old_trace = link_trace;
}
}
};
static Real SteepestDescentStep(std::vector<GaugeMat> &U,Real & alpha, GaugeMat & dmuAmu) {
GridBase *grid = U[0]._grid;
std::vector<GaugeMat> A(Nd,grid);
GaugeMat g(grid);
GaugeLinkToLieAlgebraField(U,A);
ExpiAlphaDmuAmu(A,g,alpha,dmuAmu);
Real vol = grid->gSites();
Real trG = TensorRemove(sum(trace(g))).real()/vol/Nc;
SU<Nc>::GaugeTransform(U,g);
return trG;
}
static Real FourierAccelSteepestDescentStep(std::vector<GaugeMat> &U,Real & alpha, GaugeMat & dmuAmu) {
GridBase *grid = U[0]._grid;
Real vol = grid->gSites();
FFT theFFT((GridCartesian *)grid);
LatticeComplex Fp(grid);
LatticeComplex psq(grid); psq=zero;
LatticeComplex pmu(grid);
LatticeComplex one(grid); one = Complex(1.0,0.0);
GaugeMat g(grid);
GaugeMat dmuAmu_p(grid);
std::vector<GaugeMat> A(Nd,grid);
GaugeLinkToLieAlgebraField(U,A);
DmuAmu(A,dmuAmu);
theFFT.FFT_all_dim(dmuAmu_p,dmuAmu,FFT::forward);
//////////////////////////////////
// Work out Fp = psq_max/ psq...
//////////////////////////////////
std::vector<int> latt_size = grid->GlobalDimensions();
std::vector<int> coor(grid->_ndimension,0);
for(int mu=0;mu<Nd;mu++) {
Real TwoPiL = M_PI * 2.0/ latt_size[mu];
LatticeCoordinate(pmu,mu);
pmu = TwoPiL * pmu ;
psq = psq + 4.0*sin(pmu*0.5)*sin(pmu*0.5);
}
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;
theFFT.FFT_all_dim(dmuAmu,dmuAmu_p,FFT::backward);
GaugeMat ciadmam(grid);
Complex cialpha(0.0,-alpha);
ciadmam = dmuAmu*cialpha;
SU<Nc>::taExp(ciadmam,g);
Real trG = TensorRemove(sum(trace(g))).real()/vol/Nc;
SU<Nc>::GaugeTransform(U,g);
return trG;
}
static void ExpiAlphaDmuAmu(const std::vector<GaugeMat> &A,GaugeMat &g,Real & alpha, GaugeMat &dmuAmu) {
GridBase *grid = g._grid;
Complex cialpha(0.0,-alpha);
GaugeMat ciadmam(grid);
DmuAmu(A,dmuAmu);
ciadmam = dmuAmu*cialpha;
SU<Nc>::taExp(ciadmam,g);
}
};
}
}
#endif

3
lib/qcd/utils/SUn.h
View File

@ -716,8 +716,7 @@ template<typename GaugeField,typename GaugeMat>
for (int a = 0; a < AdjointDimension; a++) {
generator(a, Ta);
auto tmp = - 2.0 * (trace(timesI(Ta) * in)) * scale;// 2.0 for the normalization of the trace in the fundamental rep
pokeColour(h_out, tmp, a);
pokeColour(h_out, - 2.0 * (trace(timesI(Ta) * in)) * scale, a);
}
}

3
lib/qcd/utils/Utils.h
View File

@ -12,7 +12,4 @@
#include <Grid/qcd/utils/SUnAdjoint.h>
#include <Grid/qcd/utils/SUnTwoIndex.h>
#endif

34
lib/qcd/utils/WilsonLoops.h
View File

@ -73,7 +73,7 @@ public:
//////////////////////////////////////////////////
// trace of directed plaquette oriented in mu,nu plane
//////////////////////////////////////////////////
static void traceDirPlaquette(LatticeComplex &plaq,
static void traceDirPlaquette(ComplexField &plaq,
const std::vector<GaugeMat> &U, const int mu,
const int nu) {
GaugeMat sp(U[0]._grid);
@ -83,9 +83,9 @@ public:
//////////////////////////////////////////////////
// sum over all planes of plaquette
//////////////////////////////////////////////////
static void sitePlaquette(LatticeComplex &Plaq,
static void sitePlaquette(ComplexField &Plaq,
const std::vector<GaugeMat> &U) {
LatticeComplex sitePlaq(U[0]._grid);
ComplexField sitePlaq(U[0]._grid);
Plaq = zero;
for (int mu = 1; mu < Nd; mu++) {
for (int nu = 0; nu < mu; nu++) {
@ -104,11 +104,11 @@ public:
U[mu] = PeekIndex<LorentzIndex>(Umu, mu);
}
LatticeComplex Plaq(Umu._grid);
ComplexField Plaq(Umu._grid);
sitePlaquette(Plaq, U);
TComplex Tp = sum(Plaq);
Complex p = TensorRemove(Tp);
auto Tp = sum(Plaq);
auto p = TensorRemove(Tp);
return p.real();
}
@ -129,15 +129,15 @@ public:
static RealD linkTrace(const GaugeLorentz &Umu) {
std::vector<GaugeMat> U(Nd, Umu._grid);
LatticeComplex Tr(Umu._grid);
ComplexField Tr(Umu._grid);
Tr = zero;
for (int mu = 0; mu < Nd; mu++) {
U[mu] = PeekIndex<LorentzIndex>(Umu, mu);
Tr = Tr + trace(U[mu]);
}
TComplex Tp = sum(Tr);
Complex p = TensorRemove(Tp);
auto Tp = sum(Tr);
auto p = TensorRemove(Tp);
double vol = Umu._grid->gSites();
@ -355,8 +355,8 @@ static void StapleMult(GaugeMat &staple, const GaugeLorentz &Umu, int mu) {
double coeff = 8.0/(32.0*M_PI*M_PI);
LatticeComplex qfield = coeff*trace(Bx*Ex + By*Ey + Bz*Ez);
TComplex Tq = sum(qfield);
ComplexField qfield = coeff*trace(Bx*Ex + By*Ey + Bz*Ez);
auto Tq = sum(qfield);
return TensorRemove(Tq).real();
}
@ -375,16 +375,16 @@ static void StapleMult(GaugeMat &staple, const GaugeLorentz &Umu, int mu) {
adj(Gimpl::CovShiftForward(
U[nu], nu, Gimpl::CovShiftForward(U[nu], nu, U[mu])));
}
static void traceDirRectangle(LatticeComplex &rect,
static void traceDirRectangle(ComplexField &rect,
const std::vector<GaugeMat> &U, const int mu,
const int nu) {
GaugeMat sp(U[0]._grid);
dirRectangle(sp, U, mu, nu);
rect = trace(sp);
}
static void siteRectangle(LatticeComplex &Rect,
static void siteRectangle(ComplexField &Rect,
const std::vector<GaugeMat> &U) {
LatticeComplex siteRect(U[0]._grid);
ComplexField siteRect(U[0]._grid);
Rect = zero;
for (int mu = 1; mu < Nd; mu++) {
for (int nu = 0; nu < mu; nu++) {
@ -404,12 +404,12 @@ static void StapleMult(GaugeMat &staple, const GaugeLorentz &Umu, int mu) {
U[mu] = PeekIndex<LorentzIndex>(Umu, mu);
}
LatticeComplex Rect(Umu._grid);
ComplexField Rect(Umu._grid);
siteRectangle(Rect, U);
TComplex Tp = sum(Rect);
Complex p = TensorRemove(Tp);
auto Tp = sum(Rect);
auto p = TensorRemove(Tp);
return p.real();
}
//////////////////////////////////////////////////

4
lib/serialisation/Hdf5IO.cc
View File

@ -65,10 +65,12 @@ Hdf5Reader::Hdf5Reader(const std::string &fileName)
Hdf5Type<unsigned int>::type());
}
void Hdf5Reader::push(const std::string &s)
bool Hdf5Reader::push(const std::string &s)
{
group_ = group_.openGroup(s);
path_.push_back(s);
return true;
}
void Hdf5Reader::pop(void)

2
lib/serialisation/Hdf5IO.h
View File

@ -54,7 +54,7 @@ namespace Grid
public:
Hdf5Reader(const std::string &fileName);
virtual ~Hdf5Reader(void) = default;
void push(const std::string &s);
bool push(const std::string &s);
void pop(void);
template <typename U>
void readDefault(const std::string &s, U &output);

3
lib/serialisation/MacroMagic.h
View File

@ -110,11 +110,12 @@ THE SOFTWARE.
#define GRID_MACRO_MEMBER(A,B) A B;
#define GRID_MACRO_COMP_MEMBER(A,B) result = (result and (lhs. B == rhs. B));
#define GRID_MACRO_OS_WRITE_MEMBER(A,B) os<< #A <<" "#B <<" = "<< obj. B <<" ; " <<std::endl;
#define GRID_MACRO_OS_WRITE_MEMBER(A,B) os<< #A <<" " #B << " = " << obj. B << " ; " <<std::endl;
#define GRID_MACRO_READ_MEMBER(A,B) Grid::read(RD,#B,obj. B);
#define GRID_MACRO_WRITE_MEMBER(A,B) Grid::write(WR,#B,obj. B);
#define GRID_SERIALIZABLE_CLASS_MEMBERS(cname,...)\
std::string SerialisableClassName(void) {return std::string(#cname);} \
GRID_MACRO_EVAL(GRID_MACRO_MAP(GRID_MACRO_MEMBER,__VA_ARGS__))\
template <typename T>\
static inline void write(Writer<T> &WR,const std::string &s, const cname &obj){ \

58
lib/serialisation/XmlIO.cc
View File

@ -32,16 +32,21 @@ using namespace Grid;
using namespace std;
// Writer implementation ///////////////////////////////////////////////////////
XmlWriter::XmlWriter(const string &fileName)
: fileName_(fileName)
XmlWriter::XmlWriter(const string &fileName, string toplev) : fileName_(fileName)
{
node_ = doc_.append_child();
node_.set_name("grid");
if ( toplev == std::string("") ) {
node_=doc_;
} else {
node_=doc_.append_child();
node_.set_name(toplev.c_str());
}
}
XmlWriter::~XmlWriter(void)
{
doc_.save_file(fileName_.c_str(), " ");
if ( fileName_ != std::string("") ) {
doc_.save_file(fileName_.c_str(), " ");
}
}
void XmlWriter::push(const string &s)
@ -53,21 +58,44 @@ void XmlWriter::pop(void)
{
node_ = node_.parent();
}
// Reader implementation ///////////////////////////////////////////////////////
XmlReader::XmlReader(const string &fileName)
: fileName_(fileName)
std::string XmlWriter::XmlString(void)
{
pugi::xml_parse_result result = doc_.load_file(fileName_.c_str());
if ( !result )
{
std::ostringstream oss;
doc_.save(oss);
return oss.str();
}
XmlReader::XmlReader(const char *xmlstring,string toplev) : fileName_("")
{
pugi::xml_parse_result result;
result = doc_.load_string(xmlstring);
if ( !result ) {
cerr << "XML error description: " << result.description() << "\n";
cerr << "XML error offset : " << result.offset << "\n";
abort();
}
node_ = doc_.child("grid");
if ( toplev == std::string("") ) {
node_ = doc_;
} else {
node_ = doc_.child(toplev.c_str());
}
}
// Reader implementation ///////////////////////////////////////////////////////
XmlReader::XmlReader(const string &fileName,string toplev) : fileName_(fileName)
{
pugi::xml_parse_result result;
result = doc_.load_file(fileName_.c_str());
if ( !result ) {
cerr << "XML error description: " << result.description() << "\n";
cerr << "XML error offset : " << result.offset << "\n";
abort();
}
if ( toplev == std::string("") ) {
node_ = doc_;
} else {
node_ = doc_.child(toplev.c_str());
}
}
bool XmlReader::push(const string &s)

11
lib/serialisation/XmlIO.h
View File

@ -44,10 +44,9 @@ namespace Grid
{
class XmlWriter: public Writer<XmlWriter>
{
{
public:
XmlWriter(const std::string &fileName);
XmlWriter(const std::string &fileName,std::string toplev = std::string("grid") );
virtual ~XmlWriter(void);
void push(const std::string &s);
void pop(void);
@ -55,6 +54,7 @@ namespace Grid
void writeDefault(const std::string &s, const U &x);
template <typename U>
void writeDefault(const std::string &s, const std::vector<U> &x);
std::string XmlString(void);
private:
pugi::xml_document doc_;
pugi::xml_node node_;
@ -64,7 +64,8 @@ namespace Grid
class XmlReader: public Reader<XmlReader>
{
public:
XmlReader(const std::string &fileName);
XmlReader(const char *xmlstring,std::string toplev = std::string("grid") );
XmlReader(const std::string &fileName,std::string toplev = std::string("grid") );
virtual ~XmlReader(void) = default;
bool push(const std::string &s);
void pop(void);
@ -118,7 +119,7 @@ namespace Grid
std::string buf;
readDefault(s, buf);
std::cout << s << " " << buf << std::endl;
// std::cout << s << " " << buf << std::endl;
fromString(output, buf);
}

25
lib/simd/Grid_avx.h
View File

@ -701,9 +701,28 @@ namespace Optimization {
//Integer Reduce
template<>
inline Integer Reduce<Integer, __m256i>::operator()(__m256i in){
// FIXME unimplemented
printf("Reduce : Missing integer implementation -> FIX\n");
assert(0);
__m128i ret;
#if defined (AVX2)
// AVX2 horizontal adds within upper and lower halves of register; use
// SSE to add upper and lower halves for result.
__m256i v1, v2;
__m128i u1, u2;
v1 = _mm256_hadd_epi32(in, in);
v2 = _mm256_hadd_epi32(v1, v1);
u1 = _mm256_castsi256_si128(v2); // upper half
u2 = _mm256_extracti128_si256(v2, 1); // lower half
ret = _mm_add_epi32(u1, u2);
#else
// No AVX horizontal add; extract upper and lower halves of register & use
// SSE intrinsics.
__m128i u1, u2, u3;
u1 = _mm256_extractf128_si256(in, 0); // upper half
u2 = _mm256_extractf128_si256(in, 1); // lower half
u3 = _mm_add_epi32(u1, u2);
u1 = _mm_hadd_epi32(u3, u3);
ret = _mm_hadd_epi32(u1, u1);
#endif
return _mm_cvtsi128_si32(ret);
}
}

22
lib/simd/Grid_avx512.h
View File

@ -543,6 +543,24 @@ namespace Optimization {
u512d conv; conv.v = v1;
return conv.f[0];
}
//Integer Reduce
template<>
inline Integer Reduce<Integer, __m512i>::operator()(__m512i in){
// No full vector reduce, use AVX to add upper and lower halves of register
// and perform AVX reduction.
__m256i v1, v2, v3;
__m128i u1, u2, ret;
v1 = _mm512_castsi512_si256(in); // upper half
v2 = _mm512_extracti32x8_epi32(in, 1); // lower half
v3 = _mm256_add_epi32(v1, v2);
v1 = _mm256_hadd_epi32(v3, v3);
v2 = _mm256_hadd_epi32(v1, v1);
u1 = _mm256_castsi256_si128(v2) // upper half
u2 = _mm256_extracti128_si256(v2, 1); // lower half
ret = _mm_add_epi32(u1, u2);
return _mm_cvtsi128_si32(ret);
}
#else
//Complex float Reduce
template<>
@ -570,9 +588,7 @@ namespace Optimization {
//Integer Reduce
template<>
inline Integer Reduce<Integer, __m512i>::operator()(__m512i in){
// FIXME unimplemented
printf("Reduce : Missing integer implementation -> FIX\n");
assert(0);
return _mm512_reduce_add_epi32(in);
}
#endif

4
lib/simd/Grid_imci.h
View File

@ -401,9 +401,7 @@ namespace Optimization {
//Integer Reduce
template<>
inline Integer Reduce<Integer, __m512i>::operator()(__m512i in){
// FIXME unimplemented
printf("Reduce : Missing integer implementation -> FIX\n");
assert(0);
return _mm512_reduce_add_epi32(in);
}

1
lib/simd/Grid_neon.h
View File

@ -596,3 +596,4 @@ namespace Optimization {
typedef Optimization::TimesI TimesISIMD;
}

90
lib/simd/Grid_qpx.h
View File

@ -374,6 +374,84 @@ namespace Optimization {
// Complex float
FLOAT_WRAP_2(operator(), inline)
};
#define USE_FP16
struct PrecisionChange {
static inline vech StoH (const vector4float &a, const vector4float &b) {
vech ret;
std::cout << GridLogError << "QPX single to half precision conversion not yet supported." << std::endl;
assert(0);
return ret;
}
static inline void HtoS (vech h, vector4float &sa, vector4float &sb) {
std::cout << GridLogError << "QPX half to single precision conversion not yet supported." << std::endl;
assert(0);
}
static inline vector4float DtoS (vector4double a, vector4double b) {
vector4float ret;
std::cout << GridLogError << "QPX double to single precision conversion not yet supported." << std::endl;
assert(0);
return ret;
}
static inline void StoD (vector4float s, vector4double &a, vector4double &b) {
std::cout << GridLogError << "QPX single to double precision conversion not yet supported." << std::endl;
assert(0);
}
static inline vech DtoH (vector4double a, vector4double b,
vector4double c, vector4double d) {
vech ret;
std::cout << GridLogError << "QPX double to half precision conversion not yet supported." << std::endl;
assert(0);
return ret;
}
static inline void HtoD (vech h, vector4double &a, vector4double &b,
vector4double &c, vector4double &d) {
std::cout << GridLogError << "QPX half to double precision conversion not yet supported." << std::endl;
assert(0);
}
};
//////////////////////////////////////////////
// Exchange support
#define FLOAT_WRAP_EXCHANGE(fn) \
static inline void fn(vector4float &out1, vector4float &out2, \
vector4float in1, vector4float in2) \
{ \
vector4double out1d, out2d, in1d, in2d; \
in1d = Vset()(in1); \
in2d = Vset()(in2); \
fn(out1d, out2d, in1d, in2d); \
Vstore()(out1d, out1); \
Vstore()(out2d, out2); \
}
struct Exchange{
// double precision
static inline void Exchange0(vector4double &out1, vector4double &out2,
vector4double in1, vector4double in2) {
out1 = vec_perm(in1, in2, vec_gpci(0145));
out2 = vec_perm(in1, in2, vec_gpci(02367));
}
static inline void Exchange1(vector4double &out1, vector4double &out2,
vector4double in1, vector4double in2) {
out1 = vec_perm(in1, in2, vec_gpci(0426));
out2 = vec_perm(in1, in2, vec_gpci(01537));
}
static inline void Exchange2(vector4double &out1, vector4double &out2,
vector4double in1, vector4double in2) {
assert(0);
}
static inline void Exchange3(vector4double &out1, vector4double &out2,
vector4double in1, vector4double in2) {
assert(0);
}
// single precision
FLOAT_WRAP_EXCHANGE(Exchange0);
FLOAT_WRAP_EXCHANGE(Exchange1);
FLOAT_WRAP_EXCHANGE(Exchange2);
FLOAT_WRAP_EXCHANGE(Exchange3);
};
struct Permute{
//Complex double
@ -497,15 +575,19 @@ namespace Optimization {
//Integer Reduce
template<>
inline Integer Reduce<Integer, int>::operator()(int in){
// FIXME unimplemented
printf("Reduce : Missing integer implementation -> FIX\n");
assert(0);
inline Integer Reduce<Integer, veci>::operator()(veci in){
Integer a = 0;
for (unsigned int i = 0; i < W<Integer>::r; ++i)
{
a += in.v[i];
}
return a;
}
}
////////////////////////////////////////////////////////////////////////////////
// Here assign types
typedef Optimization::vech SIMD_Htype; // Half precision type
typedef Optimization::vector4float SIMD_Ftype; // Single precision type
typedef vector4double SIMD_Dtype; // Double precision type
typedef Optimization::veci SIMD_Itype; // Integer type

6
lib/simd/Grid_sse4.h
View File

@ -570,9 +570,9 @@ namespace Optimization {
//Integer Reduce
template<>
inline Integer Reduce<Integer, __m128i>::operator()(__m128i in){
// FIXME unimplemented
printf("Reduce : Missing integer implementation -> FIX\n");
assert(0);
__m128i v1 = _mm_hadd_epi32(in, in);
__m128i v2 = _mm_hadd_epi32(v1, v1);
return _mm_cvtsi128_si32(v2);
}
}

2
lib/simd/Grid_vector_types.h
View File

@ -53,7 +53,7 @@ directory
#if defined IMCI
#include "Grid_imci.h"
#endif
#ifdef NEONv8
#ifdef NEONV8
#include "Grid_neon.h"
#endif
#if defined QPX

7
lib/simd/Grid_vector_unops.h
View File

@ -179,13 +179,6 @@ inline Grid_simd<S, V> div(const Grid_simd<S, V> &r, Integer y) {
////////////////////////////////////////////////////////////////////////////
// Allows us to assign into **conformable** real vectors from complex
////////////////////////////////////////////////////////////////////////////
// template < class S, class V >
// inline auto ComplexRemove(const Grid_simd<S,V> &c) ->
// Grid_simd<Grid_simd<S,V>::Real,V> {
// Grid_simd<Grid_simd<S,V>::Real,V> ret;
// ret.v = c.v;
// return ret;
// }
template <class scalar>
struct AndFunctor {
scalar operator()(const scalar &x, const scalar &y) const { return x & y; }

30
lib/stencil/Lebesgue.cc
View File

@ -32,8 +32,11 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
namespace Grid {
int LebesgueOrder::UseLebesgueOrder;
#ifdef KNL
std::vector<int> LebesgueOrder::Block({8,2,2,2});
#else
std::vector<int> LebesgueOrder::Block({2,2,2,2});
#endif
LebesgueOrder::IndexInteger LebesgueOrder::alignup(IndexInteger n){
n--; // 1000 0011 --> 1000 0010
n |= n >> 1; // 1000 0010 | 0100 0001 = 1100 0011
@ -51,8 +54,31 @@ LebesgueOrder::LebesgueOrder(GridBase *_grid)
if ( Block[0]==0) ZGraph();
else if ( Block[1]==0) NoBlocking();
else CartesianBlocking();
}
if (0) {
std::cout << "Thread Interleaving"<<std::endl;
ThreadInterleave();
}
}
void LebesgueOrder::ThreadInterleave(void)
{
std::vector<IndexInteger> reorder = _LebesgueReorder;
std::vector<IndexInteger> throrder;
int vol = _LebesgueReorder.size();
int threads = GridThread::GetThreads();
int blockbits=3;
int blocklen = 8;
int msk = 0x7;
for(int t=0;t<threads;t++){
for(int ss=0;ss<vol;ss++){
if ( ( ss >> blockbits) % threads == t ) {
throrder.push_back(reorder[ss]);
}
}
}
_LebesgueReorder = throrder;
}
void LebesgueOrder::NoBlocking(void)
{
std::cout<<GridLogDebug<<"Lexicographic : no cache blocking"<<std::endl;

2
lib/stencil/Lebesgue.h
View File

@ -70,6 +70,8 @@ namespace Grid {
std::vector<IndexInteger> & xi,
std::vector<IndexInteger> &dims);
void ThreadInterleave(void);
private:
std::vector<IndexInteger> _LebesgueReorder;

147
lib/stencil/Stencil.h
View File

@ -176,6 +176,9 @@ class CartesianStencil { // Stencil runs along coordinate axes only; NO diagonal
// Timing info; ugly; possibly temporary
/////////////////////////////////////////
double commtime;
double mpi3synctime;
double mpi3synctime_g;
double shmmergetime;
double gathertime;
double gathermtime;
double halogtime;
@ -185,6 +188,10 @@ class CartesianStencil { // Stencil runs along coordinate axes only; NO diagonal
double splicetime;
double nosplicetime;
double calls;
std::vector<double> comm_bytes_thr;
std::vector<double> comm_time_thr;
std::vector<double> comm_enter_thr;
std::vector<double> comm_leave_thr;
////////////////////////////////////////
// Stencil query
@ -248,35 +255,120 @@ class CartesianStencil { // Stencil runs along coordinate axes only; NO diagonal
//////////////////////////////////////////
// Comms packet queue for asynch thread
//////////////////////////////////////////
void CommunicateThreaded()
{
#ifdef GRID_OMP
// must be called in parallel region
int mythread = omp_get_thread_num();
int nthreads = CartesianCommunicator::nCommThreads;
#else
int mythread = 0;
int nthreads = 1;
#endif
if (nthreads == -1) nthreads = 1;
if (mythread < nthreads) {
comm_enter_thr[mythread] = usecond();
for (int i = mythread; i < Packets.size(); i += nthreads) {
uint64_t bytes = _grid->StencilSendToRecvFrom(Packets[i].send_buf,
Packets[i].to_rank,
Packets[i].recv_buf,
Packets[i].from_rank,
Packets[i].bytes,i);
comm_bytes_thr[mythread] += bytes;
}
comm_leave_thr[mythread]= usecond();
comm_time_thr[mythread] += comm_leave_thr[mythread] - comm_enter_thr[mythread];
}
}
void CollateThreads(void)
{
int nthreads = CartesianCommunicator::nCommThreads;
double first=0.0;
double last =0.0;
for(int t=0;t<nthreads;t++) {
double t0 = comm_enter_thr[t];
double t1 = comm_leave_thr[t];
comms_bytes+=comm_bytes_thr[t];
comm_enter_thr[t] = 0.0;
comm_leave_thr[t] = 0.0;
comm_time_thr[t] = 0.0;
comm_bytes_thr[t]=0;
if ( first == 0.0 ) first = t0; // first is t0
if ( (t0 > 0.0) && ( t0 < first ) ) first = t0; // min time seen
if ( t1 > last ) last = t1; // max time seen
}
commtime+= last-first;
}
void CommunicateBegin(std::vector<std::vector<CommsRequest_t> > &reqs)
{
reqs.resize(Packets.size());
commtime-=usecond();
for(int i=0;i<Packets.size();i++){
comms_bytes+=_grid->StencilSendToRecvFromBegin(reqs[i],
Packets[i].send_buf,
Packets[i].to_rank,
Packets[i].recv_buf,
Packets[i].from_rank,
Packets[i].bytes);
Packets[i].send_buf,
Packets[i].to_rank,
Packets[i].recv_buf,
Packets[i].from_rank,
Packets[i].bytes,i);
}
}
void CommunicateComplete(std::vector<std::vector<CommsRequest_t> > &reqs)
{
for(int i=0;i<Packets.size();i++){
_grid->StencilSendToRecvFromComplete(reqs[i]);
_grid->StencilSendToRecvFromComplete(reqs[i],i);
}
commtime+=usecond();
}
void Communicate(void)
{
#ifdef GRID_OMP
#pragma omp parallel
{
// must be called in parallel region
int mythread = omp_get_thread_num();
int maxthreads= omp_get_max_threads();
int nthreads = CartesianCommunicator::nCommThreads;
assert(nthreads <= maxthreads);
if (nthreads == -1) nthreads = 1;
#else
int mythread = 0;
int nthreads = 1;
#endif
if (mythread < nthreads) {
for (int i = mythread; i < Packets.size(); i += nthreads) {
double start = usecond();
comm_bytes_thr[mythread] += _grid->StencilSendToRecvFrom(Packets[i].send_buf,
Packets[i].to_rank,
Packets[i].recv_buf,
Packets[i].from_rank,
Packets[i].bytes,i);
comm_time_thr[mythread] += usecond() - start;
}
}
#ifdef GRID_OMP
}
#endif
}
template<class compressor> void HaloExchange(const Lattice<vobj> &source,compressor &compress)
{
std::vector<std::vector<CommsRequest_t> > reqs;
Prepare();
HaloGather(source,compress);
CommunicateBegin(reqs);
CommunicateComplete(reqs);
// Concurrent
//CommunicateBegin(reqs);
//CommunicateComplete(reqs);
// Sequential, possibly threaded
Communicate();
CommsMergeSHM(compress);
CommsMerge(compress);
}
@ -285,7 +377,7 @@ class CartesianStencil { // Stencil runs along coordinate axes only; NO diagonal
{
int dimension = _directions[point];
int displacement = _distances[point];
int fd = _grid->_fdimensions[dimension];
int rd = _grid->_rdimensions[dimension];
@ -337,7 +429,9 @@ class CartesianStencil { // Stencil runs along coordinate axes only; NO diagonal
template<class compressor>
void HaloGather(const Lattice<vobj> &source,compressor &compress)
{
mpi3synctime_g-=usecond();
_grid->StencilBarrier();// Synch shared memory on a single nodes
mpi3synctime_g+=usecond();
// conformable(source._grid,_grid);
assert(source._grid==_grid);
@ -397,8 +491,12 @@ class CartesianStencil { // Stencil runs along coordinate axes only; NO diagonal
CommsMerge(decompress,Mergers,Decompressions);
}
template<class decompressor> void CommsMergeSHM(decompressor decompress) {
mpi3synctime-=usecond();
_grid->StencilBarrier();// Synch shared memory on a single nodes
mpi3synctime+=usecond();
shmmergetime-=usecond();
CommsMerge(decompress,MergersSHM,DecompressionsSHM);
shmmergetime+=usecond();
}
template<class decompressor>
@ -442,7 +540,12 @@ class CartesianStencil { // Stencil runs along coordinate axes only; NO diagonal
int checkerboard,
const std::vector<int> &directions,
const std::vector<int> &distances)
: _permute_type(npoints), _comm_buf_size(npoints)
: _permute_type(npoints),
_comm_buf_size(npoints),
comm_bytes_thr(npoints),
comm_enter_thr(npoints),
comm_leave_thr(npoints),
comm_time_thr(npoints)
{
face_table_computed=0;
_npoints = npoints;
@ -996,6 +1099,15 @@ class CartesianStencil { // Stencil runs along coordinate axes only; NO diagonal
void ZeroCounters(void) {
gathertime = 0.;
commtime = 0.;
mpi3synctime=0.;
mpi3synctime_g=0.;
shmmergetime=0.;
for(int i=0;i<_npoints;i++){
comm_time_thr[i]=0;
comm_bytes_thr[i]=0;
comm_enter_thr[i]=0;
comm_leave_thr[i]=0;
}
halogtime = 0.;
mergetime = 0.;
decompresstime = 0.;
@ -1011,6 +1123,18 @@ class CartesianStencil { // Stencil runs along coordinate axes only; NO diagonal
#define PRINTIT(A) AVERAGE(A); std::cout << GridLogMessage << " Stencil " << #A << " "<< A/calls<<std::endl;
RealD NP = _grid->_Nprocessors;
RealD NN = _grid->NodeCount();
double t = 0;
// if comm_time_thr is set they were all done in parallel so take the max
// but add up the bytes
int threaded = 0 ;
for (int i = 0; i < 8; ++i) {
if ( comm_time_thr[i]>0.0 ) {
threaded = 1;
comms_bytes += comm_bytes_thr[i];
if (t < comm_time_thr[i]) t = comm_time_thr[i];
}
}
if (threaded) commtime += t;
_grid->GlobalSum(commtime); commtime/=NP;
if ( calls > 0. ) {
@ -1026,6 +1150,9 @@ class CartesianStencil { // Stencil runs along coordinate axes only; NO diagonal
std::cout << GridLogMessage << " Stencil " << comms_bytes/commtime/1000. << " GB/s per rank"<<std::endl;
std::cout << GridLogMessage << " Stencil " << comms_bytes/commtime/1000.*NP/NN << " GB/s per node"<<std::endl;
}
PRINTIT(mpi3synctime);
PRINTIT(mpi3synctime_g);
PRINTIT(shmmergetime);
PRINTIT(splicetime);
PRINTIT(nosplicetime);
}

4
lib/tensors/Tensor_arith_mul.h
View File

@ -98,7 +98,9 @@ template<class rtype,class vtype,class mtype,int N>
strong_inline void mult(iVector<rtype,N> * __restrict__ ret,
const iVector<vtype,N> * __restrict__ rhs,
const iScalar<mtype> * __restrict__ lhs){
mult(ret,lhs,rhs);
for(int c1=0;c1<N;c1++){
mult(&ret->_internal[c1],&rhs->_internal[c1],&lhs->_internal);
}
}

9
lib/tensors/Tensor_class.h
View File

@ -156,11 +156,18 @@ class iScalar {
// 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) {
strong_inline iScalar<vtype> operator=(T arg) {
_internal = arg;
return *this;
}
// Convert elements
template <class ttype>
strong_inline iScalar<vtype> operator=(iScalar<ttype> &&arg) {
_internal = arg._internal;
return *this;
}
friend std::ostream &operator<<(std::ostream &stream,const iScalar<vtype> &o) {
stream << "S {" << o._internal << "}";
return stream;

7
lib/tensors/Tensor_exp.h
View File

@ -80,8 +80,11 @@ template<class vtype, int N> inline iVector<vtype, N> Exponentiate(const iVector
mat iQ2 = arg*arg*alpha*alpha;
mat iQ3 = arg*iQ2*alpha;
// sign in c0 from the conventions on the Ta
c0 = -imag( trace(iQ3) ) * one_over_three;
c1 = -real( trace(iQ2) ) * one_over_two;
scalar imQ3, reQ2;
imQ3 = imag( trace(iQ3) );
reQ2 = real( trace(iQ2) );
c0 = -imQ3 * one_over_three;
c1 = -reQ2 * one_over_two;
// Cayley Hamilton checks to machine precision, tested
tmp = c1 * one_over_three;

60
lib/tensors/Tensor_index.h
View File

@ -47,6 +47,28 @@ template<int Level>
class TensorIndexRecursion {
public:
////////////////////////////////////////////////////
// Type Queries
////////////////////////////////////////////////////
template<class vtype> static inline int indexRank(const iScalar<vtype> tmp) { return TensorIndexRecursion<Level-1>::indexRank(tmp._internal); }
template<class vtype,int N> static inline int indexRank(const iVector<vtype,N> tmp){ return TensorIndexRecursion<Level-1>::indexRank(tmp._internal[0]); }
template<class vtype,int N> static inline int indexRank(const iMatrix<vtype,N> tmp){ return TensorIndexRecursion<Level-1>::indexRank(tmp._internal[0][0]); }
template<class vtype> static inline int isScalar(const iScalar<vtype> tmp) { return TensorIndexRecursion<Level-1>::isScalar(tmp._internal); }
template<class vtype,int N> static inline int isScalar(const iVector<vtype,N> tmp){ return TensorIndexRecursion<Level-1>::isScalar(tmp._internal[0]); }
template<class vtype,int N> static inline int isScalar(const iMatrix<vtype,N> tmp){ return TensorIndexRecursion<Level-1>::isScalar(tmp._internal[0][0]); }
template<class vtype> static inline int isVector(const iScalar<vtype> tmp) { return TensorIndexRecursion<Level-1>::isVector(tmp._internal); }
template<class vtype,int N> static inline int isVector(const iVector<vtype,N> tmp){ return TensorIndexRecursion<Level-1>::isVector(tmp._internal[0]); }
template<class vtype,int N> static inline int isVector(const iMatrix<vtype,N> tmp){ return TensorIndexRecursion<Level-1>::isVector(tmp._internal[0][0]); }
template<class vtype> static inline int isMatrix(const iScalar<vtype> tmp) { return TensorIndexRecursion<Level-1>::isMatrix(tmp._internal); }
template<class vtype,int N> static inline int isMatrix(const iVector<vtype,N> tmp){ return TensorIndexRecursion<Level-1>::isMatrix(tmp._internal[0]); }
template<class vtype,int N> static inline int isMatrix(const iMatrix<vtype,N> tmp){ return TensorIndexRecursion<Level-1>::isMatrix(tmp._internal[0][0]); }
////////////////////////////////////////////////////
// Trace
////////////////////////////////////////////////////
template<class vtype>
static auto traceIndex(const iScalar<vtype> arg) -> iScalar<decltype(TensorIndexRecursion<Level-1>::traceIndex(arg._internal))>
{
@ -215,6 +237,24 @@ class TensorIndexRecursion {
template<>
class TensorIndexRecursion<0> {
public:
////////////////////////////////////////////////////
// Type Queries
////////////////////////////////////////////////////
template<class vtype> static inline int indexRank(const iScalar<vtype> tmp) { return 1; }
template<class vtype,int N> static inline int indexRank(const iVector<vtype,N> tmp){ return N; }
template<class vtype,int N> static inline int indexRank(const iMatrix<vtype,N> tmp){ return N; }
template<class vtype> static inline int isScalar(const iScalar<vtype> tmp) { return true;}
template<class vtype,int N> static inline int isScalar(const iVector<vtype,N> tmp){ return false;}
template<class vtype,int N> static inline int isScalar(const iMatrix<vtype,N> tmp){ return false;}
template<class vtype> static inline int isVector(const iScalar<vtype> tmp) { return false;}
template<class vtype,int N> static inline int isVector(const iVector<vtype,N> tmp){ return true;}
template<class vtype,int N> static inline int isVector(const iMatrix<vtype,N> tmp){ return false;}
template<class vtype> static inline int isMatrix(const iScalar<vtype> tmp) { return false;}
template<class vtype,int N> static inline int isMatrix(const iVector<vtype,N> tmp){ return false;}
template<class vtype,int N> static inline int isMatrix(const iMatrix<vtype,N> tmp){ return true;}
/////////////////////////////////////////
// Ends recursion for trace (scalar/vector/matrix)
@ -302,6 +342,26 @@ class TensorIndexRecursion<0> {
////////////////////////////////////////////////////////////////////////////////////////////////////////
// External wrappers
////////////////////////////////////////////////////////////////////////////////////////////////////////
template<int Level,class vtype> inline int indexRank(void)
{
vtype tmp;
return TensorIndexRecursion<Level>::indexRank(tmp);
}
template<int Level,class vtype> inline int isScalar(void)
{
vtype tmp;
return TensorIndexRecursion<Level>::isScalar(tmp);
}
template<int Level,class vtype> inline int isVector(void)
{
vtype tmp;
return TensorIndexRecursion<Level>::isVector(tmp);
}
template<int Level,class vtype> inline int isMatrix(void)
{
vtype tmp;
return TensorIndexRecursion<Level>::isMatrix(tmp);
}
template<int Level,class vtype> inline auto traceIndex (const vtype &arg) -> RemoveCRV(TensorIndexRecursion<Level>::traceIndex(arg))
{

4
lib/tensors/Tensor_traits.h
View File

@ -281,8 +281,8 @@ namespace Grid {
template<typename T>
class getPrecision{
public:
typedef typename getVectorType<T>::type vector_obj; //get the vector_obj (i.e. a grid Tensor) if its a Lattice<vobj>, do nothing otherwise (i.e. if fundamental or grid Tensor)
//get the vector_obj (i.e. a grid Tensor) if its a Lattice<vobj>, do nothing otherwise (i.e. if fundamental or grid Tensor)
typedef typename getVectorType<T>::type vector_obj;
typedef typename GridTypeMapper<vector_obj>::scalar_type scalar_type; //get the associated scalar type. Works on fundamental and tensor types
typedef typename GridTypeMapper<scalar_type>::Realified real_scalar_type; //remove any std::complex wrapper, should get us to the fundamental type

24
lib/util/Init.cc
View File

@ -222,6 +222,11 @@ void Grid_init(int *argc,char ***argv)
CartesianCommunicator::MAX_MPI_SHM_BYTES = MB*1024*1024;
}
if( GridCmdOptionExists(*argv,*argv+*argc,"--shm-hugepages") ){
CartesianCommunicator::Hugepages = 1;
}
if( GridCmdOptionExists(*argv,*argv+*argc,"--debug-signals") ){
Grid_debug_handler_init();
}
@ -304,6 +309,7 @@ void Grid_init(int *argc,char ***argv)
std::cout<<GridLogMessage<<" --threads n : default number of OMP threads"<<std::endl;
std::cout<<GridLogMessage<<" --grid n.n.n.n : default Grid size"<<std::endl;
std::cout<<GridLogMessage<<" --shm M : allocate M megabytes of shared memory for comms"<<std::endl;
std::cout<<GridLogMessage<<" --shm-hugepages : use explicit huge pages in mmap call "<<std::endl;
std::cout<<GridLogMessage<<std::endl;
std::cout<<GridLogMessage<<"Verbose and debug:"<<std::endl;
std::cout<<GridLogMessage<<std::endl;
@ -317,7 +323,7 @@ void Grid_init(int *argc,char ***argv)
std::cout<<GridLogMessage<<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<<" --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;
std::cout<<GridLogMessage<<" --dslash-unroll : Wilson kernel for Nc=3"<<std::endl;
@ -356,10 +362,15 @@ void Grid_init(int *argc,char ***argv)
if( GridCmdOptionExists(*argv,*argv+*argc,"--comms-sequential") ){
CartesianCommunicator::SetCommunicatorPolicy(CartesianCommunicator::CommunicatorPolicySequential);
}
if( GridCmdOptionExists(*argv,*argv+*argc,"--lebesgue") ){
LebesgueOrder::UseLebesgueOrder=1;
}
CartesianCommunicator::nCommThreads = -1;
if( GridCmdOptionExists(*argv,*argv+*argc,"--comms-threads") ){
arg= GridCmdOptionPayload(*argv,*argv+*argc,"--comms-threads");
GridCmdOptionInt(arg,CartesianCommunicator::nCommThreads);
}
if( GridCmdOptionExists(*argv,*argv+*argc,"--cacheblocking") ){
arg= GridCmdOptionPayload(*argv,*argv+*argc,"--cacheblocking");
GridCmdOptionIntVector(arg,LebesgueOrder::Block);
@ -374,10 +385,13 @@ void Grid_init(int *argc,char ***argv)
Grid_default_latt,
Grid_default_mpi);
std::cout << GridLogDebug << "Requesting "<< CartesianCommunicator::MAX_MPI_SHM_BYTES <<" byte stencil comms buffers "<<std::endl;
std::cout << GridLogMessage << "Requesting "<< CartesianCommunicator::MAX_MPI_SHM_BYTES <<" byte stencil comms buffers "<<std::endl;
if ( CartesianCommunicator::Hugepages) {
std::cout << GridLogMessage << "Mapped stencil comms buffers as MAP_HUGETLB "<<std::endl;
}
if( GridCmdOptionExists(*argv,*argv+*argc,"--decomposition") ){
std::cout<<GridLogMessage<<"Grid Decomposition\n";
std::cout<<GridLogMessage<<"Grid Default Decomposition patterns\n";
std::cout<<GridLogMessage<<"\tOpenMP threads : "<<GridThread::GetThreads()<<std::endl;
std::cout<<GridLogMessage<<"\tMPI tasks : "<<GridCmdVectorIntToString(GridDefaultMpi())<<std::endl;
std::cout<<GridLogMessage<<"\tvRealF : "<<sizeof(vRealF)*8 <<"bits ; " <<GridCmdVectorIntToString(GridDefaultSimd(4,vRealF::Nsimd()))<<std::endl;
@ -393,7 +407,7 @@ void Grid_init(int *argc,char ***argv)
void Grid_finalize(void)
{
#if defined (GRID_COMMS_MPI) || defined (GRID_COMMS_MPI3)
#if defined (GRID_COMMS_MPI) || defined (GRID_COMMS_MPI3) || defined (GRID_COMMS_MPIT)
MPI_Finalize();
Grid_unquiesce_nodes();
#endif