static int
bicgilu_cl_qop_generic( QOP_FermionLinksWilson *qop_links,
QOP_invert_arg_t *qop_invert_arg,
QOP_resid_arg_t ***qop_resid_arg,
MYREAL *kappas[], int nkappa[],
QOP_DiracFermion **qop_sol[],
QOP_DiracFermion *qop_src[],
int nsrc,
int *final_restart,
Real *final_rsq_ptr )
{
int isrc, ikappa;
int iters;
QOP_info_t info;
if(nsrc == 1 && nkappa[0] == 1)
QOP_wilson_invert( &info, qop_links, qop_invert_arg, qop_resid_arg[0][0],
kappas[0][0], qop_sol[0][0], qop_src[0] );
else
QOP_wilson_invert_multi( &info, qop_links, qop_invert_arg, qop_resid_arg,
kappas, nkappa, qop_sol, qop_src, nsrc );
/* For now we return the largest value and total iterations */
*final_rsq_ptr = 0;
*final_restart = 0;
iters = 0;
for(isrc = 0; isrc < nsrc; isrc++)
for(ikappa = 0; ikappa < nkappa[isrc]; ikappa++){
if(*final_rsq_ptr < qop_resid_arg[isrc][ikappa]->final_rsq)
*final_rsq_ptr = qop_resid_arg[isrc][ikappa]->final_rsq;
if(*final_restart < qop_resid_arg[isrc][ikappa]->final_restart)
*final_restart = qop_resid_arg[isrc][ikappa]->final_restart;
iters += qop_resid_arg[isrc][ikappa]->final_iter;
#ifdef CG_DEBUG
if(nsrc > 1 || nkappa[isrc] > 1)
node0_printf("CONGRAD5(src %d,kappa %d): iters = %d resid = %e\n",
isrc, ikappa,
qop_resid_arg[isrc][ikappa]->final_iter,
qop_resid_arg[isrc][ikappa]->final_rsq);
#endif
}
#ifdef CGTIME
node0_printf("CGTIME: time = %e (wilson_qop %s) ",
info.final_sec,qop_prec[QOP_Precision-1]);
for(isrc = 0; isrc < nsrc; isrc++)
node0_printf("nkappa[%d] = %d iters = %d ",
isrc,nkappa[isrc],qop_resid_arg[isrc][0]->final_iter);
node0_printf("mflops = %e\n", info.final_flop/(1.0e6*info.final_sec) );
fflush(stdout);
#endif
return iters;
}
double
bench_inv(QOP_info_t *info, QOP_invert_arg_t *inv_arg,
QOP_resid_arg_t *res_arg, QDP_DiracFermion *out,
QDP_DiracFermion *in)
{
static QLA_Real r2s=-1, r2;
double sec=0, flop=0, mf=0;
int i, iter=0;
QOP_DiracFermion *qopout, *qopin;
QDP_D_eq_zero(out, QDP_all);
qopout = QOP_create_D_from_qdp(out);
qopin = QOP_create_D_from_qdp(in);
for(i=0; i<=nit; i++) {
QMP_barrier();
QOP_wilson_invert(info, flw, inv_arg, res_arg, kappa, qopout, qopin);
QMP_barrier();
printf("%i\t%i\t%g\t%i\n", i, res_arg->final_iter, info->final_sec, (int)info->final_flop);
if(i>0) {
iter += res_arg->final_iter;
sec += info->final_sec;
flop += info->final_flop;
//mf += info->final_flop/(1e6*info->final_sec);
}
}
QOP_destroy_D(qopout);
QOP_destroy_D(qopin);
QDP_r_eq_norm2_D(&r2, out, QDP_even);
if(r2s<0) r2s = r2;
if(fabs(1-r2/r2s)>1e-3) {
printf0("first norm = %g this norn = %g\n", r2s, r2);
}
mf = 1;
QMP_sum_double(&mf);
QMP_sum_double(&sec);
QMP_sum_double(&flop);
res_arg->final_iter = iter/nit;
info->final_sec = sec/(mf*nit);
info->final_flop = flop/(mf*nit);
mf = info->final_flop/(1e6*info->final_sec);
return mf;
}
