/*--------------------------------------------------------------------*/
void clear_v_field(su3_vector *v){
int i;
site *s;
FORALLSITES(i,s){
clearvec( &v[i] );
}
/* Thin the random source */
static void
thin_source(su3_vector *src, int thinning, int ex, int ey, int ez, int et){
site *s;
int i;
FORALLSITES(i,s) {
if(s->x % thinning != ex || s->y % thinning != ey || s->z % thinning != ez || s->t % thinning != et){
clearvec(src+i);
}
}
}
void accum_gauge_hit(int gauge_dir,int parity)
{
/* Accumulates sums and differences of link matrices for determining optimum */
/* hit for gauge fixing */
/* Differences are kept in diffmat and the diagonal elements of the sums */
/* in sumvec */
register int j;
register su3_matrix *m1,*m2;
register int dir,i;
register site *s;
/* Clear sumvec and diffmat */
FORSOMEPARITY(i,s,parity)
{
clear_su3mat(&diffmatp[i]);
clearvec(&sumvecp[i]);
}
int quark_renorm( void ) {
register int i, dir;
register site *s;
int j, cgn;
Real mass_x2, finalrsq;
Real pix, piy, piz, pit;
Real sin_pmu, q_mu, prop_a, prop_b, z_fac, m_func, ftmp = 0;
Real r1, r2, r3;
int pmu, px, py, pz, pt;
int pxn, pyn, pzn, ptn;
int currentnode;
int j1, jm2, k, dirs[4];
msg_tag *mtag[2];
int j_mass;
su3_vector **psim = NULL;
int xi, j2, j3, j4, parity;
int multiflag;
FILE *fp_mom_ks[MAX_NUM_MASS]; /* for writing mom propagator files */
char filename[50];
int prec = PRECISION; /* Make internal precision for CG the same as
the prevailing precision */
pix = 2.*PI / (Real)nx;
piy = 2.*PI / (Real)ny;
piz = 2.*PI / (Real)nz;
pit = 2.*PI / (Real)nt;
cgn = 0;
if( num_mass == 1){
multiflag = 0;
}
else{
multiflag = 1;
psim = (su3_vector **)malloc(num_mass*sizeof(su3_vector *));
for(j=0; j<num_mass; j++){
psim[j] = (su3_vector *)malloc(sites_on_node*sizeof(su3_vector));
}
}
/* Open the momentum propagator files */
if(this_node == 0){
for(j=0; j<num_mass; j++){
sprintf(filename,"mom_pt_prop.m_%d",j);
fp_mom_ks[j] = fopen(filename, "ab");
if(fp_mom_ks[j] == NULL){
printf("quark_renorm: Node %d can't open file %s, error %d\n",
this_node,filename,errno);fflush(stdout);
terminate(1);
}
}
}
rephase( ON ); /* Turn staggered phases on */
/* Create fat and long links */
load_ferm_links(&fn_links, &ks_act_paths);
/* Loop over the 16 source points */
for(xi=0; xi<16; xi++){
/* Initialize color trace of the propagator */
FORALLSITES(i,s){
for(j=0; j<num_mass; j++){
s->trace_prop[j].real = 0.0;
s->trace_prop[j].imag = 0.0;
}
}
j1 = xi%2;
k = xi/2;
j2 = k%2;
k /= 2;
j3 = k%2;
k /= 2;
j4 = k%2;
parity = (j1+j2+j3+j4)%2;
/* dirs[XUP] = j1;
dirs[YUP] = j2;
dirs[ZUP] = j3;
dirs[TUP] = j4; */
/* Loop over colors of source vector */
for(j=0; j<3; j++){
/* initialize the source in phi */
FORALLSITES(i,s){
clearvec( &(s->phi));
}
/* Point source at site xi in the hypercube at origin */
if( node_number(j1,j2,j3,j4) == this_node ){
i=node_index(j1,j2,j3,j4);
lattice[i].phi.c[j].real = -1.0;
}
if( multiflag == 0){
for(j_mass=0; j_mass<num_mass; j_mass++){
FORALLSITES(i,s){
clearvec( &(s->xxx1));
}
if(parity == 0){
/* do a C.G. (source in phi, result in xxx1) */
cgn += ks_congrad( F_OFFSET(phi), F_OFFSET(xxx1),
mass[j_mass], niter, nrestart,
rsqprop, PRECISION,
EVEN, &finalrsq, &fn_links);
/* Multiply by -Madjoint */
dslash_site( F_OFFSET(xxx1), F_OFFSET(ttt), ODD,
&fn_links);
mass_x2 = 2.*mass[j_mass];
FOREVENSITES(i,s){
scalar_mult_su3_vector( &(s->xxx1), -mass_x2,
&(s->ttt));
}
}
else{
// -----------------------------------------------------------------
// Matrix--vector operation
// Applies either the operator (sign = 1) or its adjoint (sign = -1)
// Adjoint is simply overall negative sign...
void fermion_op(vector *src, vector *dest, int sign) {
register int i;
register site *s;
int dir, a, b, c, d, par, L[NDIMS] = {nx, ny, nz, nt};
Real tr, halfG = 0.5 * G, m_ov_G, vev[DIMF][DIMF];
vector tvec, tvec_dir, tvec_opp;
msg_tag *tag[2 * NDIMS];
// Quick sanity check
if (sign != 1 && sign != -1) {
node0_printf("Error: incorrect sign in fermion_op: %d\n", sign);
terminate(1);
}
// Ignore site_mass if G = 0 to avoid dividing by zero
// Could be made more robust, but unlikely to matter
if (G == 0.0)
m_ov_G = 0.0;
else
m_ov_G = 2.0 * site_mass / G;
for (a = 0; a < DIMF; a++) {
for (b = 0; b < DIMF; b++)
vev[a][b] = 0.0;
}
vev[0][1] = m_ov_G;
vev[2][3] = m_ov_G;
vev[1][0] = -m_ov_G;
vev[3][2] = -m_ov_G;
// Start gathers for kinetic term
FORALLUPDIR(dir) {
if (L[dir] <= 1) // Will be skipped below
continue;
tag[dir] = start_gather_field(src, sizeof(vector), dir,
EVENANDODD, gen_pt[dir]);
tag[OPP_DIR(dir)] = start_gather_field(src, sizeof(vector), OPP_DIR(dir),
EVENANDODD, gen_pt[OPP_DIR(dir)]);
}
// Compute scalar term as gathers run
// Initialize dest = 0.5G * (sigma + 2m / G) * src
// Add SO(4)-breaking 'site mass' term with same structure as sigma
FORALLSITES(i, s) {
clearvec(&(dest[i]));
if (stagger == -1 || lattice[i].parity == EVEN)
par = 1;
else // Both stagger == 1 and lattice[i].parity == ODD
par = -1;
for (a = 0; a < DIMF; a++) {
for (b = a + 1; b < DIMF; b++) {
tr = s->sigma.e[as_index[a][b]] + par * vev[a][b];
for (c = 0; c < DIMF; c++) {
for (d = c + 1; d < DIMF; d++) {
tr += perm[a][b][c][d] * (s->sigma.e[as_index[c][d]]
+ par * vev[c][d]);
}
} // No half since not double-counting
dest[i].c[a] += tr * src[i].c[b];
dest[i].c[b] -= tr * src[i].c[a];
}
}
scalar_mult_vec(&(dest[i]), halfG, &(dest[i]));
}
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;
