static void ks_multicg_reverse_field( /* Return value is number of iterations taken */
su3_vector *src, /* source vector (type su3_vector) */
su3_vector **psim, /* solution vectors */
ks_param *ksp, /* KS parametes, including the offsets */
int num_offsets, /* number of offsets */
quark_invert_control *qic,
imp_ferm_links_t *fn /* Storage for fermion links */
)
{
char myname[] = "ks_multicg_reverse_field";
/* Site su3_vector's resid, cg_p and ttt are used as temporaies */
register int i;
register site *s;
int iteration; /* counter for iterations */
int num_offsets_now; /* number of offsets still being worked on */
double c1, c2, rsq, oldrsq, pkp; /* pkp = cg_p.K.cg_p */
double source_norm; /* squared magnitude of source vector */
double rsqstop; /* stopping residual normalized by source norm */
int l_parity=0; /* parity we are currently doing */
int l_otherparity=0; /* the other parity */
#ifdef FN
msg_tag *tags1[16], *tags2[16]; /* tags for gathers to parity and opposite */
#endif
int special_started; /* 1 if dslash_special has been called */
int j, j_low;
Real *shifts, mass_low, msq_xm4;
double *zeta_i, *zeta_im1, *zeta_ip1;
double *beta_i, *beta_im1, *alpha;
// su3_vector **pm; /* vectors not involved in gathers */
// Switch indices
su3_vector **psim_rev; su3_vector *psim_space;
su3_vector **pm_rev; su3_vector *pm_space;
/* Unpack structure */
/* We don't restart this algorithm, so we adopt the convention of
taking the product here */
int niter = qic->max*qic->nrestart;
Real rsqmin = qic->resid * qic->resid; /* desired squared residual -
normalized as sqrt(r*r)/sqrt(src_e*src_e) */
int parity = qic->parity; /* EVEN, ODD */
/* Timing */
#ifdef CGTIME
double dtimec;
#endif
double nflop;
qic->final_iters = 0;
qic->final_restart = 0;
//#if FERM_ACTION == HISQ
// fn->hl.current_X_set = 0;
// restore_fn_links(fn);
//#endif
if( num_offsets==0 )return;
if(fn == NULL){
printf("%s(%d): Called with NULL fn\n", myname, this_node);
terminate(1);
}
// Switch indices
psim_rev = (su3_vector **)malloc( sizeof(su3_vector *)*sites_on_node );
psim_space = (su3_vector *)malloc( sizeof(su3_vector)*sites_on_node*num_offsets );
pm_rev = (su3_vector **)malloc( sizeof(su3_vector *)*sites_on_node );
pm_space = (su3_vector *)malloc( sizeof(su3_vector)*sites_on_node*num_offsets );
if( psim_space == NULL || pm_space == NULL){printf("%s: NO ROOM!\n",myname); exit(0); }
for( i=0; i<sites_on_node; i++ ){
psim_rev[i] = &(psim_space[num_offsets*i]);
pm_rev[i] = &(pm_space[num_offsets*i]);
for( j=0; j<num_offsets; j++){
psim_rev[i][j] = psim[j][i];
}
}
/* debug */
#ifdef CGTIME
dtimec = -dclock();
#endif
nflop = 1205 + 15*num_offsets;
if(parity==EVENANDODD)nflop *=2;
special_started = 0;
/* if we want both parities, we will do even first. */
switch(parity){
case(EVEN): l_parity=EVEN; l_otherparity=ODD; break;
case(ODD): l_parity=ODD; l_otherparity=EVEN; break;
case(EVENANDODD): l_parity=EVEN; l_otherparity=ODD; break;
}
shifts = (Real *)malloc(num_offsets*sizeof(Real));
zeta_i = (double *)malloc(num_offsets*sizeof(double));
zeta_im1 = (double *)malloc(num_offsets*sizeof(double));
zeta_ip1 = (double *)malloc(num_offsets*sizeof(double));
beta_i = (double *)malloc(num_offsets*sizeof(double));
beta_im1 = (double *)malloc(num_offsets*sizeof(double));
alpha = (double *)malloc(num_offsets*sizeof(double));
//pm = (su3_vector **)malloc(num_offsets*sizeof(su3_vector *));
mass_low = 1.0e+20;
j_low = -1;
for(j=0;j<num_offsets;j++){
shifts[j] = ksp[j].offset;
if (ksp[j].offset < mass_low){
mass_low = ksp[j].offset;
j_low = j;
}
}
for(j=0;j<num_offsets;j++) if(j!=j_low){
//pm[j] = (su3_vector *)malloc(sites_on_node*sizeof(su3_vector));
shifts[j] -= shifts[j_low];
}
msq_xm4 = -shifts[j_low];
iteration = 0;
#define PAD 0
/* now we can allocate temporary variables and copy then */
/* PAD may be used to avoid cache thrashing */
if(first_multicongrad) {
ttt = (su3_vector *) malloc((sites_on_node+PAD)*sizeof(su3_vector));
cg_p = (su3_vector *) malloc((sites_on_node+PAD)*sizeof(su3_vector));
resid = (su3_vector *) malloc((sites_on_node+PAD)*sizeof(su3_vector));
first_multicongrad = 0;
}
#ifdef CGTIME
dtimec = -dclock();
#endif
/* initialization process */
start:
#ifdef FN
if(special_started==1) { /* clean up gathers */
cleanup_gathers(tags1, tags2);
special_started = 0;
}
#endif
num_offsets_now = num_offsets;
source_norm = 0.0;
FORSOMEPARITY(i,s,l_parity){
source_norm += (double) magsq_su3vec( src+i );
su3vec_copy( src+i, &(resid[i]));
su3vec_copy(&(resid[i]), &(cg_p[i]));
clearvec(&(psim_rev[i][j_low]));
for(j=0;j<num_offsets;j++) if(j!=j_low){
clearvec(&(psim_rev[i][j]));
su3vec_copy(&(resid[i]), &(pm_rev[i][j]));
}
} END_LOOP;
Until we have EO versions of dslash_field, we require FN
#endif
#include "generic_ks_includes.h" /* definitions files and prototypes */
#include "../include/dslash_ks_redefine.h" /* Actually not used, yet */
#include "../include/loopend.h"
/*#define CGTIME*/
int ks_congrad_two_src( /* Return value is number of iterations taken */
field_offset src1, /* source vector (type su3_vector) */
field_offset src2,
field_offset dest1, /* solution vectors */
field_offset dest2,
Real mass1,
Real mass2,
int niter, /* maximal number of CG interations */
int nrestart, /* maximal number of CG restarts */
Real rsqmin, /* desired residue squared */
int prec, /* internal precision for the inversion (ignored) */
int parity, /* parity to be worked on */
Real *final_rsq_ptr, /* final residue squared */
ferm_links_t *fn /* Storage for fermion links */
)
{
/* Site su3_vector's resid, cg_p and ttt are used as temporaries */
register int i;
register site *s;
int iteration; /* counter for iterations */
double c1, c2, rsq, oldrsq, pkp; /* pkp = cg_p.K.cg_p */
double source_norm,source_norm1; /* squared magnitude of source vector */
double rsqstop; /* stopping residual normalized by source norm */
int l_parity; /* parity we are currently doing */
int l_otherparity; /* the other parity */
msg_tag *tags1[16], *tags2[16]; /* tags for gathers to parity and opposite */
int special_started; /* 1 if dslash_special has been called */
int j, jud, jstrange;
double shift, msq_xm4;
double zeta_i[2], zeta_im1[2], zeta_ip1[2];
double beta_i[2], beta_im1[2], alpha[2];
int first;
su3_vector *temp;
su3_vector *destvec1;
su3_vector *destvec2;
su3_vector *pm_strange;
su3_vector *init_guess;
su3_vector *common_source;
su3_vector *ttt;
su3_vector *cg_p;
su3_vector *resid;
/* Timing */
#ifdef CGTIME
double dtimed,dtimec;
#endif
double nflop;
/* debug */
#ifdef CGTIME
dtimec = -dclock();
#endif
first = 0;
nflop = 1187; /* THIS LOOKS WRONG - DT */
if(parity==EVENANDODD)nflop *=2;
special_started = 0;
/* if we want both parities, we will do even first. */
switch(parity){
case(EVEN): l_parity=EVEN; l_otherparity=ODD; break;
case(ODD): l_parity=ODD; l_otherparity=EVEN; break;
case(EVENANDODD): l_parity=EVEN; l_otherparity=ODD; break;
}
jud = 0;
jstrange = 1;
temp=(su3_vector *)malloc(sites_on_node*sizeof(su3_vector));
destvec1=(su3_vector *)malloc(sites_on_node*sizeof(su3_vector));
destvec2=(su3_vector *)malloc(sites_on_node*sizeof(su3_vector));
ttt=(su3_vector *)malloc(sites_on_node*sizeof(su3_vector));
cg_p=(su3_vector *)malloc(sites_on_node*sizeof(su3_vector));
resid=(su3_vector *)malloc(sites_on_node*sizeof(su3_vector));
pm_strange = (su3_vector *)malloc(sites_on_node*sizeof(su3_vector));
init_guess = (su3_vector *)malloc(sites_on_node*sizeof(su3_vector) );
common_source = (su3_vector *)malloc(sites_on_node*sizeof(su3_vector) );
shift = 4.0*( mass2*mass2 - mass1*mass1);
msq_xm4 = -4.0*mass1*mass1;
iteration = 0;
#ifdef CGTIME
dtimec = -dclock();
#endif
/* initialization process */
start:
#ifdef FN
if(special_started==1) { /* clean up gathers */
cleanup_gathers(tags1, tags2);
special_started = 0;
}
#endif
/*This loop calculates init_guess = (phi2 - phi1)/shift = X0 , which
* is used to equalize the two sources*/
FORSOMEPARITY(i,s,l_parity){
sub_su3_vector( (su3_vector *)F_PT(s,src2),(su3_vector *)F_PT(s,src1),&temp[i] );
scalar_mult_su3_vector(&temp[i],1.0/shift,&init_guess[i] );
}END_LOOP
