/* the "full" operators */
void Q_pm_psi_prec(spinor * const l, spinor * const k)
{
spinorPrecWS *ws=(spinorPrecWS*)g_precWS;
_Complex double ALIGN alpha= -1.0;
if(g_prec_sequence_d_dagger_d[0]!=0.0)
{
alpha = g_prec_sequence_d_dagger_d[0];
spinorPrecondition(l,k,ws,T,L,alpha,0,1);
}
else
assign(l,k,VOLUME);
g_mu = -g_mu;
D_psi(g_spinor_field[DUM_MATRIX], l);
gamma5(l, g_spinor_field[DUM_MATRIX], VOLUME);
g_mu = -g_mu;
if(g_prec_sequence_d_dagger_d[1]!=0.0)
{
alpha = g_prec_sequence_d_dagger_d[1];
spinorPrecondition(l,l,ws,T,L,alpha,0,1);
}
D_psi(g_spinor_field[DUM_MATRIX], l);
gamma5(l, g_spinor_field[DUM_MATRIX], VOLUME);
if(g_prec_sequence_d_dagger_d[2]!=0.0)
{
alpha = g_prec_sequence_d_dagger_d[2];
spinorPrecondition(l,l,ws,T,L,alpha,0,1);
}
}
void Qov_sq_psi(spinor * const P, spinor * const S) {
Dov_psi(g_spinor_field[DUM_MATRIX], S);
gamma5(g_spinor_field[DUM_MATRIX], g_spinor_field[DUM_MATRIX], VOLUME);
Dov_psi(P,g_spinor_field[DUM_MATRIX]);
gamma5(P,P, VOLUME);
return;
}
/* This is the version for the gpu (Florian Burger)*/
void Q_minus_psi_gpu(spinor * const l, spinor * const k)
{
gamma5(k,k,VOLUME);
g_mu = -g_mu;
D_psi(l, k);
g_mu = -g_mu;
gamma5(l, l, VOLUME);
}
/* the "full" operators */
void Q_pm_psi2(spinor * const l, spinor * const k)
{
g_mu = -10.*g_mu;
D_psi(l, k);
gamma5(g_spinor_field[DUM_MATRIX], l, VOLUME);
g_mu = -g_mu/10.;
D_psi(l, g_spinor_field[DUM_MATRIX]);
gamma5(l, l, VOLUME);
}
/* This is the version for the gpu with interchanged order of gamma5 and D_psi (Florian Burger)*/
void Q_pm_psi_gpu(spinor * const l, spinor * const k)
{
gamma5(k, k, VOLUME);
g_mu = -g_mu;
D_psi(l, k);
gamma5(g_spinor_field[DUM_MATRIX], l, VOLUME);
g_mu = -g_mu;
D_psi(l, g_spinor_field[DUM_MATRIX]);
}
int invert_cloverdoublet_eo(spinor * const Even_new_s, spinor * const Odd_new_s,
spinor * const Even_new_c, spinor * const Odd_new_c,
spinor * const Even_s, spinor * const Odd_s,
spinor * const Even_c, spinor * const Odd_c,
const double precision, const int max_iter,
const int solver_flag, const int rel_prec) {
int iter = 0;
/* here comes the inversion using even/odd preconditioning */
if(g_proc_id == 0) {printf("# Using even/odd preconditioning!\n"); fflush(stdout);}
Msw_ee_inv_ndpsi(Even_new_s, Even_new_c,
Even_s, Even_c);
Hopping_Matrix(OE, g_spinor_field[DUM_DERI], Even_new_s);
Hopping_Matrix(OE, g_spinor_field[DUM_DERI+1], Even_new_c);
/* The sign is plus, since in Hopping_Matrix */
/* the minus is missing */
assign_mul_add_r(g_spinor_field[DUM_DERI], +1., Odd_s, VOLUME/2);
assign_mul_add_r(g_spinor_field[DUM_DERI+1], +1., Odd_c, VOLUME/2);
/* Do the inversion with the preconditioned */
/* matrix to get the odd sites */
/* Here we invert the hermitean operator squared */
if(g_proc_id == 0) {
printf("# Using CG for TMWILSON flavour doublet!\n");
fflush(stdout);
}
gamma5(g_spinor_field[DUM_DERI], g_spinor_field[DUM_DERI], VOLUME/2);
gamma5(g_spinor_field[DUM_DERI+1], g_spinor_field[DUM_DERI+1], VOLUME/2);
iter = cg_her_nd(Odd_new_s, Odd_new_c, g_spinor_field[DUM_DERI], g_spinor_field[DUM_DERI+1],
max_iter, precision, rel_prec,
VOLUME/2, &Qsw_pm_ndpsi);
Qsw_dagger_ndpsi(Odd_new_s, Odd_new_c,
Odd_new_s, Odd_new_c);
/* Reconstruct the even sites */
Hopping_Matrix(EO, g_spinor_field[DUM_DERI], Odd_new_s);
Hopping_Matrix(EO, g_spinor_field[DUM_DERI+1], Odd_new_c);
Msw_ee_inv_ndpsi(g_spinor_field[DUM_DERI+2], g_spinor_field[DUM_DERI+3],
g_spinor_field[DUM_DERI], g_spinor_field[DUM_DERI+1]);
/* The sign is plus, since in Hopping_Matrix */
/* the minus is missing */
assign_add_mul_r(Even_new_s, g_spinor_field[DUM_DERI+2], +1., VOLUME/2);
assign_add_mul_r(Even_new_c, g_spinor_field[DUM_DERI+3], +1., VOLUME/2);
return(iter);
}
void Q_full(spinor * const Even_new, spinor * const Odd_new,
spinor * const Even, spinor * const Odd) {
/* Even sites */
Hopping_Matrix(EO, g_spinor_field[DUM_DERI], Odd);
assign_mul_one_pm_imu(Even_new, Even, 1., VOLUME/2);
assign_add_mul_r(Even_new, g_spinor_field[DUM_DERI], -1., VOLUME/2);
gamma5(Even_new, Even_new, VOLUME/2);
/* Odd sites */
Hopping_Matrix(OE, g_spinor_field[DUM_DERI], Even);
assign_mul_one_pm_imu(Odd_new, Odd, 1., VOLUME/2);
assign_add_mul_r(Odd_new, g_spinor_field[DUM_DERI], -1., VOLUME/2);
gamma5(Odd_new, Odd_new, VOLUME/2);
}
void Qtm_pm_sym_psi(spinor * const l, spinor * const k){
/* Q_{-} */
Hopping_Matrix(EO, g_spinor_field[DUM_MATRIX+1], k);
mul_one_pm_imu_inv(g_spinor_field[DUM_MATRIX+1], -1., VOLUME/2);
Hopping_Matrix(OE, g_spinor_field[DUM_MATRIX], g_spinor_field[DUM_MATRIX+1]);
mul_one_pm_imu_inv(g_spinor_field[DUM_MATRIX], -1., VOLUME/2);
diff(l, k, g_spinor_field[DUM_MATRIX], VOLUME/2);
gamma5(l, l, VOLUME/2);
/* Q_{+} */
Hopping_Matrix(EO, l, g_spinor_field[DUM_MATRIX]);
mul_one_pm_imu_inv(l, +1., VOLUME/2);
Hopping_Matrix(OE, g_spinor_field[DUM_MATRIX+1], l);
mul_one_pm_imu_inv(g_spinor_field[DUM_MATRIX], +1., VOLUME/2);
diff(l, k, g_spinor_field[DUM_MATRIX], VOLUME/2);
gamma5(l, l, VOLUME/2);
}
/* |R>=rnorm^2 Q^2 |S> */
void norm_Q_sqr_psi(spinor * const R, spinor * const S,
const double rnorm) {
spinor *aux;
aux=lock_Dov_WS_spinor(1);
/* Term -1-s is done in D_psi! does this comment make sense for HMC? */
/* no, it doesn't, we do have to work on this */
/* here we need to set kappa = 1./(2 (-1-s) + 8) */
D_psi(R, S);
gamma5(aux, R, VOLUME);
D_psi(R, aux);
gamma5(R, R, VOLUME);
mul_r(R, rnorm*rnorm, R, VOLUME);
unlock_Dov_WS_spinor(1);
return;
}
void Qov_sq_psi_prec(spinor * const P, spinor * const S) {
spinorPrecWS *ws=(spinorPrecWS*)g_precWS;
static complex alpha={0,0};
alpha.re=ws->precExpo[0];
spinorPrecondition(P,S,ws,T,L,alpha,0,1);
Dov_psi(g_spinor_field[DUM_MATRIX], P);
gamma5(P, g_spinor_field[DUM_MATRIX], VOLUME);
alpha.re=ws->precExpo[1];
spinorPrecondition(P,P,ws,T,L,alpha,0,1);
Dov_psi(g_spinor_field[DUM_MATRIX], P);
gamma5(P, g_spinor_field[DUM_MATRIX], VOLUME);
alpha.re=ws->precExpo[2];
spinorPrecondition(P,P,ws,T,L,alpha,0,1);
return;
}
void addproj_q_invsqrt(spinor * const Q, spinor * const P, const int n, const int N) {
int j;
spinor *aux;
complex cnorm, lambda;
static double save_ev[2]={-1.,-1.};
static int * ev_sign = NULL;
if(eigenvls[0] != save_ev[0] && eigenvls[1] != save_ev[1] ) {
if(g_proc_id == 0 && g_debug_level > 1) {
printf("# Recomputing eigenvalue signs!\n");
fflush(stdout);
}
for(j = 0; j < 2; j++) {
save_ev[j] = eigenvls[j];
}
free(ev_sign);
ev_sign = (int*) malloc(n * sizeof(int));
aux=lock_Dov_WS_spinor(1);
for(j=0; j < n; j++) {
D_psi(aux, &(eigenvectors[j*evlength]));
gamma5(aux, aux, N);
lambda = scalar_prod(&(eigenvectors[j*evlength]), aux, N, 1);
if (lambda.re < 0) {
ev_sign[j] = -1;
}
else {
ev_sign[j] = 1;
}
}
unlock_Dov_WS_spinor(1);
/* free(aux_); */
}
for(j = 0; j < n; j++) {
cnorm = scalar_prod(&(eigenvectors[j*evlength]), P, N, 1);
cnorm.re *= (double)ev_sign[j];
cnorm.im *= (double)ev_sign[j];
assign_add_mul(Q, &(eigenvectors[j*evlength]), cnorm, N);
}
return;
}
void CGeoSmoother(spinor * const P, spinor * const Q, const int Ncy, const int dummy) {
spinor ** solver_field = NULL;
const int nr_sf = 5;
double musave = g_mu;
g_mu = g_mu1;
init_solver_field(&solver_field, VOLUMEPLUSRAND/2, nr_sf);
convert_lexic_to_eo(solver_field[0], solver_field[1], Q);
if(g_c_sw > 0)
assign_mul_one_sw_pm_imu_inv(EE,solver_field[2], solver_field[0], g_mu);
else
assign_mul_one_pm_imu_inv(solver_field[2], solver_field[0], +1., VOLUME/2);
Hopping_Matrix(OE, solver_field[4], solver_field[2]);
/* The sign is plus, since in Hopping_Matrix */
/* the minus is missing */
assign_mul_add_r(solver_field[4], +1., solver_field[1], VOLUME/2);
/* Do the inversion with the preconditioned */
/* matrix to get the odd sites */
gamma5(solver_field[4], solver_field[4], VOLUME/2);
if(g_c_sw > 0) {
cg_her(solver_field[3], solver_field[4], Ncy, 1.e-8, 1,
VOLUME/2, &Qsw_pm_psi);
Qsw_minus_psi(solver_field[3], solver_field[3]);
/* Reconstruct the even sites */
Hopping_Matrix(EO, solver_field[2], solver_field[3]);
assign_mul_one_sw_pm_imu_inv(EE,solver_field[4],solver_field[2], g_mu);
}
else {
cg_her(solver_field[3], solver_field[4], Ncy, 1.e-8, 1,
VOLUME/2, &Qtm_pm_psi);
Qtm_minus_psi(solver_field[3], solver_field[3]);
/* Reconstruct the even sites */
Hopping_Matrix(EO, solver_field[4], solver_field[3]);
mul_one_pm_imu_inv(solver_field[4], +1., VOLUME/2);
}
/* The sign is plus, since in Hopping_Matrix */
/* the minus is missing */
assign_add_mul_r(solver_field[2], solver_field[4], +1., VOLUME/2);
convert_eo_to_lexic(P, solver_field[2], solver_field[3]);
g_mu = musave;
finalize_solver(solver_field, nr_sf);
return;
}
void Dov_psi(spinor * const P, spinor * const S) {
double c0,c1;
spinor *s;
static int n_cheby = 0;
static int rec_coefs = 1;
ov_s = 0.5*(1./g_kappa - 8.) - 1.;
/* printf("Degree of Polynomial set to %d\n", ov_n_cheby); */
if(n_cheby != ov_n_cheby || rec_coefs) {
calculateOverlapPolynomial();
n_cheby = ov_n_cheby;
rec_coefs = 0;
}
if(dov_ws==NULL){
init_Dov_WS();
}
/* s_ = calloc(VOLUMEPLUSRAND+1, sizeof(spinor)); */
/* #if (defined SSE3 || defined SSE2 || defined SSE) */
/* s = (spinor*)(((unsigned long int)(s_)+ALIGN_BASE)&~ALIGN_BASE); */
/* #else */
/* s = s_; */
/* #endif */
s=lock_Dov_WS_spinor(0);
/* here we do with M = 1 + s */
/* M + m_ov/2 + (M - m_ov/2) \gamma_5 sign(Q(-M)) */
c0 = -(1.0 + ov_s - 0.5*m_ov);
c1 = -(1.0 + ov_s + 0.5*m_ov);
Q_over_sqrt_Q_sqr(s, ov_cheby_coef, ov_n_cheby, S, ev_qnorm, ev_minev);
gamma5(s, s, VOLUME);
assign_mul_add_mul_r(s, S, c0, c1, VOLUME);
assign(P, s, VOLUME);
/* free(s_); */
unlock_Dov_WS_spinor(0);
return;
}
int bicg(spinor * const k, spinor * const l, double eps_sq) {
int iteration;
double xxx;
xxx=0.0;
gamma5(g_spinor_field[DUM_SOLVER+1], l, VOLUME/2);
/* main loop */
for(iteration=1;iteration<=ITER_MAX_BCG;iteration++) {
/* compute the residual*/
M_psi(DUM_SOLVER,k,q_off);
xxx=diff_and_square_norm(g_spinor_field[DUM_SOLVER], g_spinor_field[DUM_SOLVER+1], VOLUME/2);
/*apply the solver step for the residual*/
M_psi(DUM_SOLVER+2,DUM_SOLVER,q_off-(2.+2.*q_off));
assign_add_mul_r(k,-1./((1.+q_off)*(1.+q_off)),g_spinor_field[DUM_SOLVER+2], VOLUME/2);
if(xxx <= eps_sq) break;
}
if(g_proc_id==0) {
sout = fopen(solvout, "a");
fprintf(sout, "%d %e %f\n",iteration,xxx, g_mu);
fclose(sout);
}
/* if the geometric series fails, redo with conjugate gradient */
if(iteration>=ITER_MAX_BCG) {
if(ITER_MAX_BCG == 0) {
iteration = 0;
}
zero_spinor_field(k,VOLUME/2);
iteration += solve_cg(k,l,q_off,eps_sq);
Q_psi(k,k,q_off);
if(ITER_MAX_BCG != 0) {
iteration -= 1000000;
}
if(g_proc_id == 0) {
sout = fopen(solvout, "a");
fprintf(sout, "%d %e\n",iteration, g_mu);
fclose(sout);
}
}
return iteration;
}
/* |R>=rnorm^n Q^n |S> */
void norm_Q_n_psi(spinor * const R, spinor * const S,
const int n, const double rnorm) {
int i;
double npar = 1.;
spinor *aux;
aux=lock_Dov_WS_spinor(1);
assign(aux, S, VOLUME);
for(i=0; i < n; i++){
D_psi(R, aux);
/* Term -1-s is done in D_psi! does this comment make sense for HMC? */
gamma5(aux, R, VOLUME);
npar *= rnorm;
}
mul_r(R, npar, aux, VOLUME);
unlock_Dov_WS_spinor(1);
return;
}
void cloverdet_derivative(const int id, hamiltonian_field_t * const hf) {
monomial * mnl = &monomial_list[id];
/* This factor 2 a missing factor 2 in trace_lambda */
(*mnl).forcefactor = 2.;
/*********************************************************************
*
* even/odd version
*
* This a term is det(\hat Q^2(\mu))
*
*********************************************************************/
g_mu = mnl->mu;
g_mu3 = mnl->rho;
boundary(mnl->kappa);
// we compute the clover term (1 + T_ee(oo)) for all sites x
sw_term(hf->gaugefield, mnl->kappa, mnl->c_sw);
// we invert it for the even sites only
sw_invert(EE, mnl->mu);
if(mnl->solver != CG && g_proc_id == 0) {
fprintf(stderr, "Bicgstab currently not implemented, using CG instead! (cloverdet_monomial.c)\n");
}
// Invert Q_{+} Q_{-}
// X_o -> DUM_DERI+1
chrono_guess(g_spinor_field[DUM_DERI+1], mnl->pf, mnl->csg_field, mnl->csg_index_array,
mnl->csg_N, mnl->csg_n, VOLUME/2, mnl->Qsq);
mnl->iter1 += cg_her(g_spinor_field[DUM_DERI+1], mnl->pf, mnl->maxiter, mnl->forceprec,
g_relative_precision_flag, VOLUME/2, mnl->Qsq);
chrono_add_solution(g_spinor_field[DUM_DERI+1], mnl->csg_field, mnl->csg_index_array,
mnl->csg_N, &mnl->csg_n, VOLUME/2);
// Y_o -> DUM_DERI
mnl->Qm(g_spinor_field[DUM_DERI], g_spinor_field[DUM_DERI+1]);
// apply Hopping Matrix M_{eo}
// to get the even sites of X_e
H_eo_sw_inv_psi(g_spinor_field[DUM_DERI+2], g_spinor_field[DUM_DERI+1], EE, -mnl->mu);
// \delta Q sandwitched by Y_o^\dagger and X_e
deriv_Sb(OE, g_spinor_field[DUM_DERI], g_spinor_field[DUM_DERI+2], hf);
// to get the even sites of Y_e
H_eo_sw_inv_psi(g_spinor_field[DUM_DERI+3], g_spinor_field[DUM_DERI], EE, mnl->mu);
// \delta Q sandwitched by Y_e^\dagger and X_o
// uses the gauge field in hf and changes the derivative fields in hf
deriv_Sb(EO, g_spinor_field[DUM_DERI+3], g_spinor_field[DUM_DERI+1], hf);
// here comes the clover term...
// computes the insertion matrices for S_eff
// result is written to swp and swm
// even/even sites sandwiched by gamma_5 Y_e and gamma_5 X_e
gamma5(g_spinor_field[DUM_DERI+2], g_spinor_field[DUM_DERI+2], VOLUME/2);
sw_spinor(EO, g_spinor_field[DUM_DERI+2], g_spinor_field[DUM_DERI+3]);
// odd/odd sites sandwiched by gamma_5 Y_o and gamma_5 X_o
gamma5(g_spinor_field[DUM_DERI], g_spinor_field[DUM_DERI], VOLUME/2);
sw_spinor(OE, g_spinor_field[DUM_DERI], g_spinor_field[DUM_DERI+1]);
// compute the contribution for the det-part
// we again compute only the insertion matrices for S_det
// the result is added to swp and swm
// even sites only!
sw_deriv(EE, mnl->mu);
// now we compute
// finally, using the insertion matrices stored in swm and swp
// we compute the terms F^{det} and F^{sw} at once
// uses the gaugefields in hf and changes the derivative field in hf
sw_all(hf, mnl->kappa, mnl->c_sw);
g_mu = g_mu1;
g_mu3 = 0.;
boundary(g_kappa);
return;
}
/* k output , l input */
int bicg(spinor * const k, spinor * const l, double eps_sq, const int rel_prec) {
double err, d1, squarenorm=0.;
complex rho0, rho1, omega, alpha, beta, nom, denom;
int iteration, N=VOLUME/2;
spinor * r, * p, * v, *hatr, * s, * t, * P, * Q;
if(ITER_MAX_BCG > 0) {
hatr = g_spinor_field[DUM_SOLVER];
r = g_spinor_field[DUM_SOLVER+1];
v = g_spinor_field[DUM_SOLVER+2];
p = g_spinor_field[DUM_SOLVER+3];
s = g_spinor_field[DUM_SOLVER+4];
t = g_spinor_field[DUM_SOLVER+5];
P = k;
Q = l;
squarenorm = square_norm(Q, VOLUME/2, 1);
Mtm_plus_psi(r, P);
gamma5(g_spinor_field[DUM_SOLVER], l, VOLUME/2);
diff(p, hatr, r, N);
assign(r, p, N);
assign(hatr, p, N);
rho0 = scalar_prod(hatr, r, N, 1);
for(iteration = 0; iteration < ITER_MAX_BCG; iteration++){
err = square_norm(r, N, 1);
if(g_proc_id == g_stdio_proc && g_debug_level > 1) {
printf("BiCGstab: iterations: %d res^2 %e\n", iteration, err);
fflush(stdout);
}
if (((err <= eps_sq) && (rel_prec == 0))
|| ((err <= eps_sq*squarenorm) && (rel_prec == 1))){
break;
}
Mtm_plus_psi(v, p);
denom = scalar_prod(hatr, v, N, 1);
_div_complex(alpha, rho0, denom);
assign(s, r, N);
assign_diff_mul(s, v, alpha, N);
Mtm_plus_psi(t, s);
omega = scalar_prod(t,s, N, 1);
d1 = square_norm(t, N, 1);
omega.re/=d1; omega.im/=d1;
assign_add_mul_add_mul(P, p, s, alpha, omega, N);
assign(r, s, N);
assign_diff_mul(r, t, omega, N);
rho1 = scalar_prod(hatr, r, N, 1);
_mult_assign_complex(nom, alpha, rho1);
_mult_assign_complex(denom, omega, rho0);
_div_complex(beta, nom, denom);
omega.re=-omega.re; omega.im=-omega.im;
assign_mul_bra_add_mul_ket_add(p, v, r, omega, beta, N);
rho0.re = rho1.re; rho0.im = rho1.im;
}
if(g_proc_id==0 && g_debug_level > 0) {
printf("BiCGstab: iterations: %d eps_sq: %1.4e\n", iteration, eps_sq);
}
}
else{
iteration = ITER_MAX_BCG;
gamma5(k, l, VOLUME/2);
}
/* if bicg fails, redo with conjugate gradient */
if(iteration>=ITER_MAX_BCG){
iteration = solve_cg(k,l,eps_sq, rel_prec);
/* Save the solution for reuse! not needed since Chronological inverter is there */
/* assign(g_spinor_field[DUM_DERI+6], k, VOLUME/2); */
Qtm_minus_psi(k, k);;
}
return iteration;
}
void Q_plus_psi(spinor * const l, spinor * const k)
{
D_psi(l, k);
gamma5(l, l, VOLUME);
}
/* Checks GW relation only by applying Dov to delta(0,0) */
void ov_check_ginsparg_wilson_relation(void) {
double norm_diff, norm_left, norm_right, norm, max_rel = 0.0, min_left = 1.0e100, min_right = 1.0e100, max_diff = 0.0, min_norm = 1.0e100;
int x, y, z, t, i;
spinor *S_left, *S_right, *S_diff;
/* get memory for the spinor fields */
S_left = ov_alloc_spinor();
S_right = ov_alloc_spinor();
S_diff = ov_alloc_spinor();
/* Create delta source at origin */
source_spinor_field(g_spinor_field[1], g_spinor_field[0], 0, 0);
convert_eo_to_lexic(g_spinor_field[2], g_spinor_field[1], g_spinor_field[0]);
/* S_right = D gamma_5 D */
Dov_psi(g_spinor_field[3], g_spinor_field[2]);
gamma5(S_left, g_spinor_field[3], VOLUME);
Dov_psi(S_right, S_left);
/* S_left = {gamma_5, D} */
gamma5(g_spinor_field[3], g_spinor_field[2], VOLUME);
Dov_psi(g_spinor_field[4], g_spinor_field[3]);
add(S_left, S_left, g_spinor_field[4], VOLUME);
/* S_diff = S_left-S_right */
diff(S_diff, S_left, S_right, VOLUME);
/* scan the whole lattice */
printf("// test of the Ginsparg-Wilson relation\n");
for(x=0; x<LX; x++)
for(y = 0; y < LY; y++)
for(z = 0; z < LZ; z++)
for(t = 0; t < T; t++) {
i = g_ipt[t][x][y][z];
_spinor_norm_sq(norm_diff, S_diff[i]);
_spinor_norm_sq(norm_left, S_left[i]);
_spinor_norm_sq(norm_right, S_right[i]);
norm_diff = sqrt(norm_diff);
norm_left = sqrt(norm_left);
norm_right = sqrt(norm_right);
norm = norm_left+norm_right;
if (norm < min_norm)
min_norm = norm;
if (norm > 0.0) {
norm = 2.*norm_diff/norm;
if (norm > max_rel)
max_rel = norm;
}
if (norm_left < min_left)
min_left = norm_left;
if (norm_right < min_right)
min_right = norm_right;
if (norm_diff > max_diff)
max_diff = norm_diff;
}
/* print results */
printf(" - maximum absoulte deviation: %.4le\n", max_diff);
printf(" - maximum relative deviation: %.4le\n", max_rel);
printf(" - minimum mean norm: %4.le\n", min_norm);
printf(" - minimum norm {gamma_5, D}: %.4le\n", min_left);
printf(" - minimum norm D gamma_5 D: %.4le\n", min_right);
/* free memory */
ov_free_spinor(S_left);
ov_free_spinor(S_right);
ov_free_spinor(S_diff);
}
void Qov_psi(spinor * const P, spinor * const S) {
Dov_psi(P, S);
gamma5(P, P, VOLUME);
return;
}
int invert_doublet_eo(spinor * const Even_new_s, spinor * const Odd_new_s,
spinor * const Even_new_c, spinor * const Odd_new_c,
spinor * const Even_s, spinor * const Odd_s,
spinor * const Even_c, spinor * const Odd_c,
const double precision, const int max_iter,
const int solver_flag, const int rel_prec) {
int iter = 0;
#ifdef HAVE_GPU
# ifdef TEMPORALGAUGE
/* initialize temporal gauge here */
int retval;
double dret1, dret2;
double plaquette1 = 0.0;
double plaquette2 = 0.0;
if (usegpu_flag) {
/* need VOLUME here (not N=VOLUME/2)*/
if ((retval = init_temporalgauge_trafo(VOLUME, g_gauge_field)) != 0 ) { // initializes the transformation matrices
if (g_proc_id == 0) printf("Error while gauge fixing to temporal gauge. Aborting...\n"); // g_tempgauge_field as a copy of g_gauge_field
exit(200);
}
/* do trafo */
plaquette1 = measure_plaquette(g_gauge_field);
apply_gtrafo(g_gauge_field, g_trafo); // transformation of the gauge field
plaquette2 = measure_plaquette(g_gauge_field);
if (g_proc_id == 0) printf("\tPlaquette before gauge fixing: %.16e\n", plaquette1/6./VOLUME);
if (g_proc_id == 0) printf("\tPlaquette after gauge fixing: %.16e\n", plaquette2/6./VOLUME);
/* do trafo to odd_s part of source */
dret1 = square_norm(Odd_s, VOLUME/2 , 1);
apply_gtrafo_spinor_odd(Odd_s, g_trafo); // odd spinor transformation, strange
dret2 = square_norm(Odd_s, VOLUME/2, 1);
if (g_proc_id == 0) printf("\tsquare norm before gauge fixing: %.16e\n", dret1);
if (g_proc_id == 0) printf("\tsquare norm after gauge fixing: %.16e\n", dret2);
/* do trafo to odd_c part of source */
dret1 = square_norm(Odd_c, VOLUME/2 , 1);
apply_gtrafo_spinor_odd(Odd_c, g_trafo); // odd spinor transformation, charm
dret2 = square_norm(Odd_c, VOLUME/2, 1);
if (g_proc_id == 0) printf("\tsquare norm before gauge fixing: %.16e\n", dret1);
if (g_proc_id == 0) printf("\tsquare norm after gauge fixing: %.16e\n", dret2);
/* do trafo to even_s part of source */
dret1 = square_norm(Even_s, VOLUME/2 , 1);
apply_gtrafo_spinor_even(Even_s, g_trafo); // even spinor transformation, strange
dret2 = square_norm(Even_s, VOLUME/2, 1);
if (g_proc_id == 0) printf("\tsquare norm before gauge fixing: %.16e\n", dret1);
if (g_proc_id == 0) printf("\tsquare norm after gauge fixing: %.16e\n", dret2);
/* do trafo to even_c part of source */
dret1 = square_norm(Even_c, VOLUME/2 , 1);
apply_gtrafo_spinor_even(Even_c, g_trafo); // even spinor transformation, charm
dret2 = square_norm(Even_c, VOLUME/2, 1);
if (g_proc_id == 0) printf("\tsquare norm before gauge fixing: %.16e\n", dret1);
if (g_proc_id == 0) printf("\tsquare norm after gauge fixing: %.16e\n", dret2);
# ifdef MPI
xchange_gauge(g_gauge_field);
# endif
}
# endif
#endif /* HAVE_GPU*/
/* here comes the inversion using even/odd preconditioning */
if(g_proc_id == 0) {printf("# Using even/odd preconditioning!\n"); fflush(stdout);}
M_ee_inv_ndpsi(Even_new_s, Even_new_c,
Even_s, Even_c,
g_mubar, g_epsbar);
Hopping_Matrix(OE, g_spinor_field[DUM_DERI], Even_new_s);
Hopping_Matrix(OE, g_spinor_field[DUM_DERI+1], Even_new_c);
/* The sign is plus, since in Hopping_Matrix */
/* the minus is missing */
assign_mul_add_r(g_spinor_field[DUM_DERI], +1., Odd_s, VOLUME/2);
assign_mul_add_r(g_spinor_field[DUM_DERI+1], +1., Odd_c, VOLUME/2);
/* Do the inversion with the preconditioned */
/* matrix to get the odd sites */
/* Here we invert the hermitean operator squared */
if(g_proc_id == 0) {
printf("# Using CG for TMWILSON flavour doublet!\n");
fflush(stdout);
}
gamma5(g_spinor_field[DUM_DERI], g_spinor_field[DUM_DERI], VOLUME/2);
gamma5(g_spinor_field[DUM_DERI+1], g_spinor_field[DUM_DERI+1], VOLUME/2);
#ifdef HAVE_GPU
if (usegpu_flag) { // GPU, mixed precision solver
# if defined(MPI) && defined(PARALLELT)
iter = mixedsolve_eo_nd(Odd_new_s, Odd_new_c, g_spinor_field[DUM_DERI], g_spinor_field[DUM_DERI+1],
max_iter, precision, rel_prec);
# elif !defined(MPI) && !defined(PARALLELT)
iter = mixedsolve_eo_nd(Odd_new_s, Odd_new_c, g_spinor_field[DUM_DERI], g_spinor_field[DUM_DERI+1],
max_iter, precision, rel_prec);
# else
printf("MPI and/or PARALLELT are not appropriately set for the GPU implementation. Aborting...\n");
exit(-1);
# endif
}
else { // CPU, conjugate gradient
iter = cg_her_nd(Odd_new_s, Odd_new_c, g_spinor_field[DUM_DERI], g_spinor_field[DUM_DERI+1],
max_iter, precision, rel_prec,
VOLUME/2, &Qtm_pm_ndpsi);
}
#else // CPU, conjugate gradient
iter = cg_her_nd(Odd_new_s, Odd_new_c, g_spinor_field[DUM_DERI], g_spinor_field[DUM_DERI+1],
max_iter, precision, rel_prec,
VOLUME/2, &Qtm_pm_ndpsi);
#endif
Qtm_dagger_ndpsi(Odd_new_s, Odd_new_c,
Odd_new_s, Odd_new_c);
/* Reconstruct the even sites */
Hopping_Matrix(EO, g_spinor_field[DUM_DERI], Odd_new_s);
Hopping_Matrix(EO, g_spinor_field[DUM_DERI+1], Odd_new_c);
M_ee_inv_ndpsi(g_spinor_field[DUM_DERI+2], g_spinor_field[DUM_DERI+3],
g_spinor_field[DUM_DERI], g_spinor_field[DUM_DERI+1],
g_mubar, g_epsbar);
/* The sign is plus, since in Hopping_Matrix */
/* the minus is missing */
assign_add_mul_r(Even_new_s, g_spinor_field[DUM_DERI+2], +1., VOLUME/2);
assign_add_mul_r(Even_new_c, g_spinor_field[DUM_DERI+3], +1., VOLUME/2);
#ifdef HAVE_GPU
/* return from temporal gauge again */
# ifdef TEMPORALGAUGE
if (usegpu_flag) {
/* undo trafo */
/* apply_inv_gtrafo(g_gauge_field, g_trafo);*/
/* copy back the saved original field located in g_tempgauge_field -> update necessary*/
plaquette1 = measure_plaquette(g_gauge_field);
copy_gauge_field(g_gauge_field, g_tempgauge_field);
g_update_gauge_copy = 1;
plaquette2 = measure_plaquette(g_gauge_field);
if (g_proc_id == 0) printf("\tPlaquette before inverse gauge fixing: %.16e\n", plaquette1/6./VOLUME);
if (g_proc_id == 0) printf("\tPlaquette after inverse gauge fixing: %.16e\n", plaquette2/6./VOLUME);
/* undo trafo to source Even_s */
dret1 = square_norm(Even_s, VOLUME/2 , 1);
apply_inv_gtrafo_spinor_even(Even_s, g_trafo);
dret2 = square_norm(Even_s, VOLUME/2, 1);
if (g_proc_id == 0) printf("\tsquare norm before gauge fixing: %.16e\n", dret1);
if (g_proc_id == 0) printf("\tsquare norm after gauge fixing: %.16e\n", dret2);
/* undo trafo to source Even_c */
dret1 = square_norm(Even_c, VOLUME/2 , 1);
apply_inv_gtrafo_spinor_even(Even_c, g_trafo);
dret2 = square_norm(Even_c, VOLUME/2, 1);
if (g_proc_id == 0) printf("\tsquare norm before gauge fixing: %.16e\n", dret1);
if (g_proc_id == 0) printf("\tsquare norm after gauge fixing: %.16e\n", dret2);
/* undo trafo to source Odd_s */
dret1 = square_norm(Odd_s, VOLUME/2 , 1);
apply_inv_gtrafo_spinor_odd(Odd_s, g_trafo);
dret2 = square_norm(Odd_s, VOLUME/2, 1);
if (g_proc_id == 0) printf("\tsquare norm before gauge fixing: %.16e\n", dret1);
if (g_proc_id == 0) printf("\tsquare norm after gauge fixing: %.16e\n", dret2);
/* undo trafo to source Odd_c */
dret1 = square_norm(Odd_c, VOLUME/2 , 1);
apply_inv_gtrafo_spinor_odd(Odd_c, g_trafo);
dret2 = square_norm(Odd_c, VOLUME/2, 1);
if (g_proc_id == 0) printf("\tsquare norm before gauge fixing: %.16e\n", dret1);
if (g_proc_id == 0) printf("\tsquare norm after gauge fixing: %.16e\n", dret2);
// Even_new_s
dret1 = square_norm(Even_new_s, VOLUME/2 , 1);
apply_inv_gtrafo_spinor_even(Even_new_s, g_trafo);
dret2 = square_norm(Even_new_s, VOLUME/2, 1);
if (g_proc_id == 0) printf("\tsquare norm before gauge fixing: %.16e\n", dret1);
if (g_proc_id == 0) printf("\tsquare norm after gauge fixing: %.16e\n", dret2);
// Even_new_c
dret1 = square_norm(Even_new_c, VOLUME/2 , 1);
apply_inv_gtrafo_spinor_even(Even_new_c, g_trafo);
dret2 = square_norm(Even_new_c, VOLUME/2, 1);
if (g_proc_id == 0) printf("\tsquare norm before gauge fixing: %.16e\n", dret1);
if (g_proc_id == 0) printf("\tsquare norm after gauge fixing: %.16e\n", dret2);
// Odd_new_s
dret1 = square_norm(Odd_new_s, VOLUME/2 , 1);
apply_inv_gtrafo_spinor_odd(Odd_new_s, g_trafo);
dret2 = square_norm(Odd_new_s, VOLUME/2, 1);
if (g_proc_id == 0) printf("\tsquare norm before gauge fixing: %.16e\n", dret1);
if (g_proc_id == 0) printf("\tsquare norm after gauge fixing: %.16e\n", dret2);
// Odd_new_c
dret1 = square_norm(Odd_new_c, VOLUME/2 , 1);
apply_inv_gtrafo_spinor_odd(Odd_new_c, g_trafo);
dret2 = square_norm(Odd_new_c, VOLUME/2, 1);
if (g_proc_id == 0) printf("\tsquare norm before gauge fixing: %.16e\n", dret1);
if (g_proc_id == 0) printf("\tsquare norm after gauge fixing: %.16e\n", dret2);
finalize_temporalgauge();
# ifdef MPI
xchange_gauge(g_gauge_field);
# endif
}
# endif
#endif
return(iter);
}
const Field Dirac_LowModeDeflation_ExactEigen::mult_dag(const Field& f)const{
return gamma5(mult(gamma5(f)));
}
int invert_clover_eo(spinor * const Even_new, spinor * const Odd_new,
spinor * const Even, spinor * const Odd,
const double precision, const int max_iter,
const int solver_flag, const int rel_prec,solver_params_t solver_params,
su3 *** gf, matrix_mult Qsq, matrix_mult Qm) {
int iter;
if(g_proc_id == 0 && g_debug_level > 0) {
printf("# Using even/odd preconditioning!\n");
fflush(stdout);
}
assign_mul_one_sw_pm_imu_inv(EE, Even_new, Even, +g_mu);
Hopping_Matrix(OE, g_spinor_field[DUM_DERI], Even_new);
/* The sign is plus, since in Hopping_Matrix */
/* the minus is missing */
assign_mul_add_r(g_spinor_field[DUM_DERI], +1., Odd, VOLUME/2);
/* Do the inversion with the preconditioned */
/* matrix to get the odd sites */
/* Here we invert the hermitean operator squared */
gamma5(g_spinor_field[DUM_DERI], g_spinor_field[DUM_DERI], VOLUME/2);
if(g_proc_id == 0) {
//printf("# Using CG!\n");
printf("# mu = %f, kappa = %f, csw = %f\n",
g_mu/2./g_kappa, g_kappa, g_c_sw);
fflush(stdout);
}
if(solver_flag == CG) {
if(g_proc_id == 0) {
printf("# Using CG!\n");
fflush(stdout);
}
iter = cg_her(Odd_new, g_spinor_field[DUM_DERI], max_iter,
precision, rel_prec,
VOLUME/2, Qsq);
Qm(Odd_new, Odd_new);
} else if(solver_flag == INCREIGCG) {
if(g_proc_id == 0) {
printf("# Using Incremental Eig-CG!\n");
fflush(stdout);
}
iter = incr_eigcg(VOLUME/2,solver_params.eigcg_nrhs, solver_params.eigcg_nrhs1, Odd_new, g_spinor_field[DUM_DERI], solver_params.eigcg_ldh, Qsq,
solver_params.eigcg_tolsq1, solver_params.eigcg_tolsq, solver_params.eigcg_restolsq , solver_params.eigcg_rand_guess_opt,
rel_prec, max_iter, solver_params.eigcg_nev, solver_params.eigcg_vmax);
Qm(Odd_new, Odd_new);
} else {
if(g_proc_id == 0) {
printf("# This solver is not available for this operator. Exisiting!\n");
fflush(stdout);
}
return 0;
}
/* Reconstruct the even sites */
Hopping_Matrix(EO, g_spinor_field[DUM_DERI], Odd_new);
clover_inv(g_spinor_field[DUM_DERI], +1, g_mu);
/* The sign is plus, since in Hopping_Matrix */
/* the minus is missing */
assign_add_mul_r(Even_new, g_spinor_field[DUM_DERI], +1., VOLUME/2);
return(iter);
}
void invert_overlap(const int op_id, const int index_start) {
operator * optr;
void (*op)(spinor*,spinor*);
static complex alpha={0,0};
spinorPrecWS *ws;
optr = &operator_list[op_id];
op=&Dov_psi;
/* here we need to (re)compute the kernel eigenvectors */
/* for new gauge fields */
if(g_proc_id == 0) {printf("# Not using even/odd preconditioning!\n"); fflush(stdout);}
convert_eo_to_lexic(g_spinor_field[DUM_DERI], optr->sr0, optr->sr1);
convert_eo_to_lexic(g_spinor_field[DUM_DERI+1], optr->prop0, optr->prop1);
if(optr->solver == 13 ){
optr->iterations = sumr(g_spinor_field[DUM_DERI+1],g_spinor_field[DUM_DERI] , optr->maxiter, optr->eps_sq);
}
else if(optr->solver == 1 /* CG */) {
gamma5(g_spinor_field[DUM_DERI+1], g_spinor_field[DUM_DERI], VOLUME);
if(use_preconditioning==1 && g_precWS!=NULL){
ws=(spinorPrecWS*)g_precWS;
printf("# Using preconditioning (which one?)!\n");
alpha.re=ws->precExpo[2];
spinorPrecondition(g_spinor_field[DUM_DERI+1],g_spinor_field[DUM_DERI+1],ws,T,L,alpha,0,1);
/* iter = cg_her(g_spinor_field[DUM_DERI], g_spinor_field[DUM_DERI+1], max_iter, precision, */
/* rel_prec, VOLUME, &Q_pm_psi_prec); */
optr->iterations = cg_her(g_spinor_field[DUM_DERI], g_spinor_field[DUM_DERI+1], optr->maxiter, optr->eps_sq,
optr->rel_prec, VOLUME, &Qov_sq_psi_prec);
alpha.re=ws->precExpo[0];
spinorPrecondition(g_spinor_field[DUM_DERI],g_spinor_field[DUM_DERI],ws,T,L,alpha,0,1);
}
else {
printf("# Not using preconditioning (which one?)!\n");
/* iter = cg_her(g_spinor_field[DUM_DERI], g_spinor_field[DUM_DERI+1], max_iter, precision, */
/* rel_prec, VOLUME, &Q_pm_psi); */
optr->iterations = cg_her(g_spinor_field[DUM_DERI], g_spinor_field[DUM_DERI+1], optr->maxiter, optr->eps_sq,
optr->rel_prec, VOLUME, &Qov_sq_psi);
}
Qov_psi(g_spinor_field[DUM_DERI+1], g_spinor_field[DUM_DERI]);
if(use_preconditioning == 1 && g_precWS!=NULL){
ws=(spinorPrecWS*)g_precWS;
alpha.re=ws->precExpo[1];
spinorPrecondition(g_spinor_field[DUM_DERI+1],g_spinor_field[DUM_DERI+1],ws,T,L,alpha,0,1);
}
}
op(g_spinor_field[4],g_spinor_field[DUM_DERI+1]);
convert_eo_to_lexic(g_spinor_field[DUM_DERI], optr->sr0, optr->sr1);
optr->reached_prec=diff_and_square_norm(g_spinor_field[4],g_spinor_field[DUM_DERI],VOLUME);
convert_lexic_to_eo(optr->prop0, optr->prop1 , g_spinor_field[DUM_DERI+1]);
return;
}
void cloverdetratio_derivative_orig(const int no, hamiltonian_field_t * const hf) {
monomial * mnl = &monomial_list[no];
/* This factor 2* a missing factor 2 in trace_lambda */
mnl->forcefactor = 1.;
/*********************************************************************
*
* this is being run in case there is even/odd preconditioning
*
* This term is det((Q^2 + \mu_1^2)/(Q^2 + \mu_2^2))
* mu1 and mu2 are set according to the monomial
*
*********************************************************************/
/* First term coming from the second field */
/* Multiply with W_+ */
g_mu = mnl->mu;
g_mu3 = mnl->rho2; //rho2
boundary(mnl->kappa);
// we compute the clover term (1 + T_ee(oo)) for all sites x
sw_term( (const su3**) hf->gaugefield, mnl->kappa, mnl->c_sw);
// we invert it for the even sites only including mu
sw_invert(EE, mnl->mu);
if(mnl->solver != CG) {
fprintf(stderr, "Bicgstab currently not implemented, using CG instead! (detratio_monomial.c)\n");
}
mnl->Qp(g_spinor_field[DUM_DERI+2], mnl->pf);
g_mu3 = mnl->rho; // rho1
/* Invert Q_{+} Q_{-} */
/* X_W -> DUM_DERI+1 */
chrono_guess(g_spinor_field[DUM_DERI+1], g_spinor_field[DUM_DERI+2], mnl->csg_field,
mnl->csg_index_array, mnl->csg_N, mnl->csg_n, VOLUME/2, mnl->Qsq);
mnl->iter1 += cg_her(g_spinor_field[DUM_DERI+1], g_spinor_field[DUM_DERI+2], mnl->maxiter,
mnl->forceprec, g_relative_precision_flag, VOLUME/2, mnl->Qsq);
chrono_add_solution(g_spinor_field[DUM_DERI+1], mnl->csg_field, mnl->csg_index_array,
mnl->csg_N, &mnl->csg_n, VOLUME/2);
/* Y_W -> DUM_DERI */
mnl->Qm(g_spinor_field[DUM_DERI], g_spinor_field[DUM_DERI+1]);
/* apply Hopping Matrix M_{eo} */
/* to get the even sites of X */
H_eo_sw_inv_psi(g_spinor_field[DUM_DERI+2], g_spinor_field[DUM_DERI+1], EE, -mnl->mu);
/* \delta Q sandwitched by Y_o^\dagger and X_e */
deriv_Sb(OE, g_spinor_field[DUM_DERI], g_spinor_field[DUM_DERI+2], hf, mnl->forcefactor);
/* to get the even sites of Y */
H_eo_sw_inv_psi(g_spinor_field[DUM_DERI+3], g_spinor_field[DUM_DERI], EE, mnl->mu);
/* \delta Q sandwitched by Y_e^\dagger and X_o */
deriv_Sb(EO, g_spinor_field[DUM_DERI+3], g_spinor_field[DUM_DERI+1], hf, mnl->forcefactor);
// here comes the clover term...
// computes the insertion matrices for S_eff
// result is written to swp and swm
// even/even sites sandwiched by gamma_5 Y_e and gamma_5 X_e
gamma5(g_spinor_field[DUM_DERI+2], g_spinor_field[DUM_DERI+2], VOLUME/2);
sw_spinor(EO, g_spinor_field[DUM_DERI+2], g_spinor_field[DUM_DERI+3]);
// odd/odd sites sandwiched by gamma_5 Y_o and gamma_5 X_o
gamma5(g_spinor_field[DUM_DERI], g_spinor_field[DUM_DERI], VOLUME/2);
sw_spinor(OE, g_spinor_field[DUM_DERI], g_spinor_field[DUM_DERI+1]);
g_mu3 = mnl->rho2; // rho2
/* Second term coming from the second field */
/* The sign is opposite!! */
mul_r(g_spinor_field[DUM_DERI], -1., mnl->pf, VOLUME/2);
/* apply Hopping Matrix M_{eo} */
/* to get the even sites of X */
H_eo_sw_inv_psi(g_spinor_field[DUM_DERI+2], g_spinor_field[DUM_DERI+1], EE, -mnl->mu);
/* \delta Q sandwitched by Y_o^\dagger and X_e */
deriv_Sb(OE, g_spinor_field[DUM_DERI], g_spinor_field[DUM_DERI+2], hf, mnl->forcefactor);
/* to get the even sites of Y */
H_eo_sw_inv_psi(g_spinor_field[DUM_DERI+3], g_spinor_field[DUM_DERI], EE, mnl->mu);
/* \delta Q sandwitched by Y_e^\dagger and X_o */
deriv_Sb(EO, g_spinor_field[DUM_DERI+3], g_spinor_field[DUM_DERI+1], hf, mnl->forcefactor);
// here comes the clover term...
// computes the insertion matrices for S_eff
// result is written to swp and swm
// even/even sites sandwiched by gamma_5 Y_e and gamma_5 X_e
gamma5(g_spinor_field[DUM_DERI+2], g_spinor_field[DUM_DERI+2], VOLUME/2);
sw_spinor(EO, g_spinor_field[DUM_DERI+2], g_spinor_field[DUM_DERI+3]);
// odd/odd sites sandwiched by gamma_5 Y_o and gamma_5 X_o
gamma5(g_spinor_field[DUM_DERI], g_spinor_field[DUM_DERI], VOLUME/2);
sw_spinor(OE, g_spinor_field[DUM_DERI], g_spinor_field[DUM_DERI+1]);
sw_all(hf, mnl->kappa*mnl->forcefactor, mnl->c_sw);
g_mu = g_mu1;
g_mu3 = 0.;
boundary(g_kappa);
return;
}
int invert_doublet_eo(spinor * const Even_new_s, spinor * const Odd_new_s,
spinor * const Even_new_c, spinor * const Odd_new_c,
spinor * const Even_s, spinor * const Odd_s,
spinor * const Even_c, spinor * const Odd_c,
const double precision, const int max_iter,
const int solver_flag, const int rel_prec,
solver_params_t solver_params, const ExternalInverter external_inverter,
const SloppyPrecision sloppy, const CompressionType compression) {
int iter = 0;
#ifdef TM_USE_QUDA
if( external_inverter==QUDA_INVERTER ) {
return invert_doublet_eo_quda( Even_new_s, Odd_new_s, Even_new_c, Odd_new_c,
Even_s, Odd_s, Even_c, Odd_c,
precision, max_iter,
solver_flag, rel_prec, 1,
sloppy, compression );
}
#endif
#ifdef HAVE_GPU
# ifdef TEMPORALGAUGE
if (usegpu_flag) {
gtrafo_eo_nd(Even_s, Odd_s, Even_c, Odd_c,
(spinor*const)NULL, (spinor*const)NULL, (spinor*const)NULL, (spinor*const)NULL,
GTRAFO_APPLY);
}
# endif
#endif /* HAVE_GPU*/
/* here comes the inversion using even/odd preconditioning */
if(g_proc_id == 0) {printf("# Using even/odd preconditioning!\n"); fflush(stdout);}
M_ee_inv_ndpsi(Even_new_s, Even_new_c,
Even_s, Even_c,
g_mubar, g_epsbar);
Hopping_Matrix(OE, g_spinor_field[DUM_DERI], Even_new_s);
Hopping_Matrix(OE, g_spinor_field[DUM_DERI+1], Even_new_c);
/* The sign is plus, since in Hopping_Matrix */
/* the minus is missing */
assign_mul_add_r(g_spinor_field[DUM_DERI], +1., Odd_s, VOLUME/2);
assign_mul_add_r(g_spinor_field[DUM_DERI+1], +1., Odd_c, VOLUME/2);
/* Do the inversion with the preconditioned */
/* matrix to get the odd sites */
/* Here we invert the hermitean operator squared */
if(g_proc_id == 0) {
printf("# Using CG for TMWILSON flavour doublet!\n");
fflush(stdout);
}
if ( external_inverter == NO_EXT_INV ){
gamma5(g_spinor_field[DUM_DERI], g_spinor_field[DUM_DERI], VOLUME/2);
gamma5(g_spinor_field[DUM_DERI+1], g_spinor_field[DUM_DERI+1], VOLUME/2);
#ifdef HAVE_GPU
if (usegpu_flag) { // GPU, mixed precision solver
# if ( defined TM_USE_MPI && defined PARALLELT )
iter = mixedsolve_eo_nd(Odd_new_s, Odd_new_c, g_spinor_field[DUM_DERI], g_spinor_field[DUM_DERI+1],
max_iter, precision, rel_prec);
# elif ( !defined TM_USE_MPI && !defined PARALLELT )
iter = mixedsolve_eo_nd(Odd_new_s, Odd_new_c, g_spinor_field[DUM_DERI], g_spinor_field[DUM_DERI+1],
max_iter, precision, rel_prec);
# else
printf("MPI and/or PARALLELT are not appropriately set for the GPU implementation. Aborting...\n");
exit(-1);
# endif
}
else { // CPU, conjugate gradient
iter = cg_her_nd(Odd_new_s, Odd_new_c, g_spinor_field[DUM_DERI], g_spinor_field[DUM_DERI+1],
max_iter, precision, rel_prec,
VOLUME/2, &Qtm_pm_ndpsi);
}
#else // CPU, conjugate gradient
if(solver_flag == RGMIXEDCG){
iter = rg_mixed_cg_her_nd(Odd_new_s, Odd_new_c, g_spinor_field[DUM_DERI], g_spinor_field[DUM_DERI+1],
solver_params, max_iter, precision, rel_prec, VOLUME/2,
&Qtm_pm_ndpsi, &Qtm_pm_ndpsi_32);
}
else {
iter = cg_her_nd(Odd_new_s, Odd_new_c, g_spinor_field[DUM_DERI], g_spinor_field[DUM_DERI+1],
max_iter, precision, rel_prec, VOLUME/2, &Qtm_pm_ndpsi);
}
#endif
Qtm_dagger_ndpsi(Odd_new_s, Odd_new_c,
Odd_new_s, Odd_new_c);
} // if(NO_EXT_INV)
#ifdef TM_USE_QPHIX
else if (external_inverter == QPHIX_INVERTER ) {
// using QPhiX, we invert M M^dagger y = b, so we don't need gamma_5 multiplications
iter = invert_eo_qphix_twoflavour(Odd_new_s, Odd_new_c, g_spinor_field[DUM_DERI], g_spinor_field[DUM_DERI+1],
max_iter, precision, solver_flag, rel_prec,
solver_params, sloppy, compression);
// and it multiplies y internally by M^dagger, returning M^{-1} b as required
}
#endif // TM_USE_QPHIX
/* Reconstruct the even sites */
Hopping_Matrix(EO, g_spinor_field[DUM_DERI], Odd_new_s);
Hopping_Matrix(EO, g_spinor_field[DUM_DERI+1], Odd_new_c);
M_ee_inv_ndpsi(g_spinor_field[DUM_DERI+2], g_spinor_field[DUM_DERI+3],
g_spinor_field[DUM_DERI], g_spinor_field[DUM_DERI+1],
g_mubar, g_epsbar);
/* The sign is plus, since in Hopping_Matrix */
/* the minus is missing */
assign_add_mul_r(Even_new_s, g_spinor_field[DUM_DERI+2], +1., VOLUME/2);
assign_add_mul_r(Even_new_c, g_spinor_field[DUM_DERI+3], +1., VOLUME/2);
#ifdef HAVE_GPU
/* return from temporal gauge again */
# ifdef TEMPORALGAUGE
if (usegpu_flag) {
gtrafo_eo_nd(Even_s, Odd_s, Even_c, Odd_c, Even_new_s, Odd_new_s, Even_new_c, Odd_new_c,
GTRAFO_REVERT);
}
# endif
#endif
return(iter);
}
int invert_cloverdoublet_eo(spinor * const Even_new_s, spinor * const Odd_new_s,
spinor * const Even_new_c, spinor * const Odd_new_c,
spinor * const Even_s, spinor * const Odd_s,
spinor * const Even_c, spinor * const Odd_c,
const double precision, const int max_iter,
const int solver_flag, const int rel_prec, solver_params_t solver_params,
const ExternalInverter external_inverter, const SloppyPrecision sloppy, const CompressionType compression) {
int iter = 0;
#ifdef TM_USE_QUDA
if( external_inverter==QUDA_INVERTER ) {
return invert_doublet_eo_quda( Even_new_s, Odd_new_s, Even_new_c, Odd_new_c,
Even_s, Odd_s, Even_c, Odd_c,
precision, max_iter,
solver_flag, rel_prec, 1,
sloppy, compression );
}
#endif
/* here comes the inversion using even/odd preconditioning */
if(g_proc_id == 0) {printf("# Using even/odd preconditioning!\n"); fflush(stdout);}
Msw_ee_inv_ndpsi(Even_new_s, Even_new_c,
Even_s, Even_c);
Hopping_Matrix(OE, g_spinor_field[DUM_DERI], Even_new_s);
Hopping_Matrix(OE, g_spinor_field[DUM_DERI+1], Even_new_c);
/* The sign is plus, since in Hopping_Matrix */
/* the minus is missing */
assign_mul_add_r(g_spinor_field[DUM_DERI], +1., Odd_s, VOLUME/2);
assign_mul_add_r(g_spinor_field[DUM_DERI+1], +1., Odd_c, VOLUME/2);
if( external_inverter == NO_EXT_INV ){
/* Do the inversion with the preconditioned */
/* matrix to get the odd sites */
/* Here we invert the hermitean operator squared */
if(g_proc_id == 0) {
printf("# Using CG for TMWILSON flavour doublet!\n");
fflush(stdout);
}
gamma5(g_spinor_field[DUM_DERI], g_spinor_field[DUM_DERI], VOLUME/2);
gamma5(g_spinor_field[DUM_DERI+1], g_spinor_field[DUM_DERI+1], VOLUME/2);
if(solver_flag == RGMIXEDCG){
iter = rg_mixed_cg_her_nd(Odd_new_s, Odd_new_c, g_spinor_field[DUM_DERI], g_spinor_field[DUM_DERI+1],
solver_params, max_iter, precision, rel_prec, VOLUME/2,
&Qsw_pm_ndpsi, &Qsw_pm_ndpsi_32);
} else {
iter = cg_her_nd(Odd_new_s, Odd_new_c, g_spinor_field[DUM_DERI], g_spinor_field[DUM_DERI+1],
max_iter, precision, rel_prec,
VOLUME/2, &Qsw_pm_ndpsi);
}
Qsw_dagger_ndpsi(Odd_new_s, Odd_new_c,
Odd_new_s, Odd_new_c);
} // if(NO_EXT_INV)
#ifdef TM_USE_QPHIX
else if (external_inverter == QPHIX_INVERTER ) {
// using QPhiX, we invert M M^dagger y = b, so we don't need gamma_5 multiplications
iter = invert_eo_qphix_twoflavour(Odd_new_s, Odd_new_c, g_spinor_field[DUM_DERI], g_spinor_field[DUM_DERI+1],
max_iter, precision, solver_flag, rel_prec,
solver_params, sloppy, compression);
// and it multiplies y internally by M^dagger, returning M^{-1} b as required
}
#endif // TM_USE_QPHIX
/* Reconstruct the even sites */
Hopping_Matrix(EO, g_spinor_field[DUM_DERI], Odd_new_s);
Hopping_Matrix(EO, g_spinor_field[DUM_DERI+1], Odd_new_c);
Msw_ee_inv_ndpsi(g_spinor_field[DUM_DERI+2], g_spinor_field[DUM_DERI+3],
g_spinor_field[DUM_DERI], g_spinor_field[DUM_DERI+1]);
/* The sign is plus, since in Hopping_Matrix */
/* the minus is missing */
assign_add_mul_r(Even_new_s, g_spinor_field[DUM_DERI+2], +1., VOLUME/2);
assign_add_mul_r(Even_new_c, g_spinor_field[DUM_DERI+3], +1., VOLUME/2);
return(iter);
}
/* Check GW relation with operator norm over the full lattice */
void ov_check_ginsparg_wilson_relation_strong(void) {
double norm_diff, norm_left, norm_right, norm, max_rel = 0.0, min_left = 1.0e100, min_right = 1.0e100, max_diff = 0.0, min_norm = 1.0e100;
int x, y, z, t, i, j, k;
spinor *S_left[4][3], *S_right[4][3], *S_diff[4][3];
if (g_debug_level>0) {
printf("// creating spinor fields and calculating {gamma_5,D} psi and a D gamma_5 D psi\n");
fflush(stdout);
}
for (i=0; i<4; i++)
for (j=0; j<3; j++) {
if (g_debug_level>1) {
printf("// spinor field: delta_dirac at %d, delta_color at %d\n", i, j);
fflush(stdout);
}
/* get memory for the spinor */
S_left[i][j] = ov_alloc_spinor();
S_right[i][j] = ov_alloc_spinor();
S_diff[i][j] = ov_alloc_spinor();
/* Create delta source at origin */
source_spinor_field(g_spinor_field[1], g_spinor_field[0], i, j);
convert_eo_to_lexic(g_spinor_field[2], g_spinor_field[1], g_spinor_field[0]);
/* S_right = D gamma_5 D psi */
Dov_psi(g_spinor_field[3], g_spinor_field[2]);
gamma5(S_left[i][j], g_spinor_field[3], VOLUME);
Dov_psi(S_right[i][j], S_left[i][j]);
/* S_left = {gamma_5, D} psi */
gamma5(g_spinor_field[3], g_spinor_field[2], VOLUME);
Dov_psi(g_spinor_field[4], g_spinor_field[3]);
add(S_left[i][j], S_left[i][j], g_spinor_field[4], VOLUME);
/* S_diff = (S_left-S_right) psi, should be zero (GW relation) */
diff(S_diff[i][j], S_left[i][j], S_right[i][j], VOLUME);
}
/* scan the whole lattice and check GW relation */
printf("// test of the Ginsparg-Wilson relation:\n");
if (g_debug_level>0)
fflush(stdout);
for(x=0; x<LX; x++)
for(y = 0; y < LY; y++)
for(z = 0; z < LZ; z++)
for(t = 0; t < T; t++) {
k = g_ipt[t][x][y][z];
norm_diff = ov_operator_colsumnorm(S_diff, k);
norm_left = ov_operator_colsumnorm(S_left, k);
norm_right = ov_operator_colsumnorm(S_right, k);
norm = (norm_left+norm_right)/2.;
if (norm < min_norm)
min_norm = norm;
if (norm > 0.0) {
norm = norm_diff/norm;
if (norm > max_rel)
max_rel = norm;
if ((norm > 1.8) && (g_debug_level)>=5) {
printf("(%d,%d,%d,%d): taxi = %d, rel = %.20le, lr = [%.4le, %.4le], diff = %.4le\n", t, x, y, z, ((x>LX/2) ? LX-x : x)+((y>LY/2) ? LY-y : y)+((z>LZ/2) ? LZ-z : z)+((t>T/2) ? T-t : t), norm, norm_left, norm_right, norm_diff);
printf("// left[0][0]:\n");
ov_print_spinor(&S_left[0][0][k]);
printf("// right[0][0]:\n");
ov_print_spinor(&S_right[0][0][k]);
printf("// diff[0][0]:\n");
ov_print_spinor(&S_diff[0][0][k]);
}
}
if (norm_left < min_left)
min_left = norm_left;
if (norm_right < min_right)
min_right = norm_right;
if (norm_diff > max_diff)
max_diff = norm_diff;
}
/* print results */
printf(" - maximum absolute deviation: %.4le\n", max_diff);
printf(" - maximum relative deviation: %.4le\n", max_rel);
printf(" - minimum mean norm: %.4le\n", min_norm);
printf(" - minimum norm {gamma_5, D}: %.4le\n", min_left);
printf(" - minimum norm D gamma_5 D: %.4le\n", min_right);
/* free memory */
for (i=0; i<4; i++)
for (j=0; j<3; j++) {
ov_free_spinor(S_left[i][j]);
ov_free_spinor(S_right[i][j]);
ov_free_spinor(S_diff[i][j]);
}
}
