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;
}
void
qopWilsonDslash(Layout *l, real *x, real *u[8], real mass, int sign,
real *y, char *sub)
{
QDP_ColorMatrix *qu[4];
QDP_DiracFermion *out, *in;
in = QDP_create_D();
out = QDP_create_D();
unpackD(l, in, y);
unpackD(l, out, x);
for(int i=0; i<4; i++) {
qu[i] = QDP_create_M();
unpackM(l, qu[i], u[2*i]);
QLA_Real two = 2;
QDP_M_eq_r_times_M(qu[i], &two, qu[i], QDP_all);
}
QOP_FermionLinksWilson *fla;
fla = QOP_wilson_create_L_from_qdp(qu, NULL);
QOP_evenodd_t eoOut=QOP_EVENODD, eoIn=QOP_EVENODD;
if(sub[0]=='e') {
eoOut = QOP_EVEN;
eoIn = QOP_ODD;
}
if(sub[0]=='o') {
eoOut = QOP_ODD;
eoIn = QOP_EVEN;
}
real kappa = 0.5/(4+mass);
QOP_wilson_dslash_qdp(NULL, fla, kappa, sign, out, in, eoOut, eoIn);
QLA_Real n2;
QDP_r_eq_norm2_D(&n2, out, QDP_all);
printf0("out2: %g\n", n2);
packD(l, x, out);
QDP_destroy_D(in);
QDP_destroy_D(out);
for(int i=0; i<4; i++) {
QDP_destroy_M(qu[i]);
}
}
int
main(int argc, char *argv[])
{
const char *msg;
int status = 1;
int mu, i;
struct QOP_CLOVER_State *clover_state;
QDP_Int *I_seed;
int i_seed;
QDP_RandomState *state;
QLA_Real plaq;
QLA_Real n[NELEMS(F)];
struct QOP_CLOVER_Gauge *c_g;
struct QOP_CLOVER_Fermion *c_f[NELEMS(F)];
double kappa;
double c_sw;
double in_eps;
int in_iter;
int log_flag;
double out_eps;
int out_iter;
int cg_status;
double run_time;
long long flops, sent, received;
/* start QDP */
QDP_initialize(&argc, &argv);
if (argc != 1 + NDIM + 6) {
printf0("ERROR: usage: %s Lx ... seed kappa c_sw iter eps log?\n",
argv[0]);
goto end;
}
for (mu = 0; mu < NDIM; mu++) {
lattice[mu] = atoi(argv[1 + mu]);
}
i_seed = atoi(argv[1 + NDIM]);
kappa = atof(argv[2 + NDIM]);
c_sw = atof(argv[3 + NDIM]);
in_iter = atoi(argv[4 + NDIM]);
in_eps = atof(argv[5 + NDIM]);
log_flag = atoi(argv[6 + NDIM]) == 0? 0: QOP_CLOVER_LOG_EVERYTHING;
/* set lattice size and create layout */
QDP_set_latsize(NDIM, lattice);
QDP_create_layout();
primary = QMP_is_primary_node();
self = QMP_get_node_number();
get_vector(network, 1, QMP_get_logical_number_of_dimensions(),
QMP_get_logical_dimensions());
get_vector(node, 0, QMP_get_logical_number_of_dimensions(),
QMP_get_logical_coordinates());
printf0("network: ");
for (i = 0; i < NDIM; i++)
printf0(" %d", network[i]);
printf0("\n");
printf0("node: ");
for (i = 0; i < NDIM; i++)
printf0(" %d", node[i]);
printf0("\n");
printf0("kappa: %20.15f\n", kappa);
printf0("c_sw: %20.15f\n", c_sw);
printf0("in_iter: %d\n", in_iter);
printf0("in_eps: %15.2e\n", in_eps);
/* allocate the gauge field */
create_Mvector(U, NELEMS(U));
create_Mvector(C, NELEMS(C));
create_Dvector(F, NELEMS(F));
I_seed = QDP_create_I();
QDP_I_eq_funci(I_seed, icoord, QDP_all);
state = QDP_create_S();
QDP_S_eq_seed_i_I(state, i_seed, I_seed, QDP_all);
for (mu = 0; mu < NELEMS(U); mu++) {
QDP_M_eq_gaussian_S(U[mu], state, QDP_all);
}
for (i = 0; i < NELEMS(F); i++) {
QDP_D_eq_gaussian_S(F[i], state, QDP_all);
}
/* build the clovers */
clover(C, U);
/* initialize CLOVER */
if (QOP_CLOVER_init(&clover_state, lattice, network, node, primary,
sublattice, NULL)) {
printf0("CLOVER_init() failed\n");
goto end;
}
if (QOP_CLOVER_import_fermion(&c_f[0], clover_state, f_reader, F[0])) {
printf0("CLOVER_import_fermion(0) failed\n");
goto end;
}
if (QOP_CLOVER_allocate_fermion(&c_f[1], clover_state)) {
printf0("CLOVER_allocate_fermion(1) failed\n");
goto end;
}
if (QOP_CLOVER_allocate_fermion(&c_f[2], clover_state)) {
printf0("CLOVER_allocate_fermion(2) failed\n");
goto end;
}
if (QOP_CLOVER_allocate_fermion(&c_f[3], clover_state)) {
printf0("CLOVER_allocate_fermion(3) failed\n");
goto end;
}
if (QOP_CLOVER_import_gauge(&c_g, clover_state, kappa, c_sw,
u_reader, c_reader, NULL)) {
printf("CLOVER_import_gauge() failed\n");
goto end;
}
QOP_CLOVER_D_operator(c_f[2], c_g, c_f[0]);
cg_status = QOP_CLOVER_D_CG(c_f[3], &out_iter, &out_eps,
c_f[2], c_g, c_f[2], in_iter, in_eps,
log_flag);
msg = QOP_CLOVER_error(clover_state);
QOP_CLOVER_performance(&run_time, &flops, &sent, &received, clover_state);
QOP_CLOVER_export_fermion(f_writer, F[3], c_f[3]);
printf0("CG status: %d\n", cg_status);
printf0("CG error message: %s\n", msg? msg: "<NONE>");
printf0("CG iter: %d\n", out_iter);
printf0("CG eps: %20.10e\n", out_eps);
printf0("CG performance: runtime %e sec\n", run_time);
printf0("CG performance: flops %.3e MFlop/s (%lld)\n",
flops * 1e-6 / run_time, flops);
printf0("CG performance: snd %.3e MB/s (%lld)\n",
sent * 1e-6 / run_time, sent);
printf0("CG performance: rcv %.3e MB (%lld)/s\n",
received * 1e-6 / run_time, received);
/* free CLOVER */
QOP_CLOVER_free_gauge(&c_g);
for (i = 0; i < NELEMS(c_f); i++)
QOP_CLOVER_free_fermion(&c_f[i]);
QOP_CLOVER_fini(&clover_state);
/* Compute plaquette */
plaq = plaquette(U);
/* field norms */
for (i = 0; i < NELEMS(F); i++)
QDP_r_eq_norm2_D(&n[i], F[i], QDP_all);
/* Display the values */
printf0("plaquette = %g\n",
plaq / (QDP_volume() * QDP_Nc * NDIM * (NDIM - 1) / 2 ));
for (i = 0; i < NELEMS(F); i++)
printf0(" |f|^2 [%d] = %20.10e\n", i, (double)(n[i]));
/* Compute and display <f[1] f[0]> */
show_dot("1|orig", F[1], F[0]);
/* Compute and display <f[1] f[3]> */
show_dot("1|solv", F[1], F[3]);
QDP_destroy_S(state);
QDP_destroy_I(I_seed);
destroy_Mvector(U, NELEMS(U));
destroy_Mvector(C, NELEMS(C));
destroy_Dvector(F, NELEMS(F));
status = 0;
end:
/* shutdown QDP */
printf0("end\n");
QDP_finalize();
return status;
}
int
main(int argc, char *argv[])
{
int status = 1;
int mu, i;
struct QOP_CLOVER_State *clover_state;
QDP_Int *I_seed;
int i_seed;
QDP_RandomState *state;
QLA_Real plaq;
QLA_Real n[NELEMS(F)];
struct QOP_CLOVER_Gauge *c_g;
struct QOP_CLOVER_Fermion *c_f[NELEMS(F)];
double kappa;
double c_sw;
/* start QDP */
QDP_initialize(&argc, &argv);
if (argc != 1 + NDIM + 3) {
printf0("ERROR: usage: %s Lx ... seed kappa c_sw\n", argv[0]);
goto end;
}
for (mu = 0; mu < NDIM; mu++) {
lattice[mu] = atoi(argv[1 + mu]);
}
i_seed = atoi(argv[1 + NDIM]);
kappa = atof(argv[2 + NDIM]);
c_sw = atof(argv[3 + NDIM]);
/* set lattice size and create layout */
QDP_set_latsize(NDIM, lattice);
QDP_create_layout();
primary = QMP_is_primary_node();
self = QMP_get_node_number();
get_vector(network, 1, QMP_get_logical_number_of_dimensions(),
QMP_get_logical_dimensions());
get_vector(node, 0, QMP_get_logical_number_of_dimensions(),
QMP_get_logical_coordinates());
printf0("network: ");
for (i = 0; i < NDIM; i++)
printf0(" %d", network[i]);
printf0("\n");
printf0("node: ");
for (i = 0; i < NDIM; i++)
printf0(" %d", node[i]);
printf0("\n");
printf0("kappa: %20.15f\n", kappa);
printf0("c_sw: %20.15f\n", c_sw);
/* allocate the gauge field */
create_Mvector(U, NELEMS(U));
create_Mvector(C, NELEMS(C));
create_Dvector(F, NELEMS(F));
I_seed = QDP_create_I();
QDP_I_eq_funci(I_seed, icoord, QDP_all);
state = QDP_create_S();
QDP_S_eq_seed_i_I(state, i_seed, I_seed, QDP_all);
for (mu = 0; mu < NELEMS(U); mu++) {
QDP_M_eq_gaussian_S(U[mu], state, QDP_all);
}
for (i = 0; i < NELEMS(F); i++) {
QDP_D_eq_gaussian_S(F[i], state, QDP_all);
}
/* build the clovers */
clover(C, U);
/* initialize CLOVER */
if (QOP_CLOVER_init(&clover_state, lattice, network, node, primary,
sublattice, NULL)) {
printf0("CLOVER_init() failed\n");
goto end;
}
if (QOP_CLOVER_import_fermion(&c_f[0], clover_state, f_reader, F[0])) {
printf0("CLOVER_import_fermion(0) failed\n");
goto end;
}
if (QOP_CLOVER_import_fermion(&c_f[1], clover_state, f_reader, F[1])) {
printf0("CLOVER_import_fermion(1) failed\n");
goto end;
}
if (QOP_CLOVER_allocate_fermion(&c_f[2], clover_state)) {
printf0("CLOVER_allocate_fermion(2) failed\n");
goto end;
}
if (QOP_CLOVER_allocate_fermion(&c_f[3], clover_state)) {
printf0("CLOVER_allocate_fermion(3) failed\n");
goto end;
}
if (QOP_CLOVER_import_gauge(&c_g, clover_state, kappa, c_sw,
u_reader, c_reader, NULL)) {
printf("CLOVER_import_gauge() failed\n");
goto end;
}
QOP_CLOVER_D_operator(c_f[2], c_g, c_f[0]);
QOP_CLOVER_export_fermion(f_writer, F[2], c_f[2]);
QOP_CLOVER_D_operator_conjugated(c_f[3], c_g, c_f[1]);
QOP_CLOVER_export_fermion(f_writer, F[3], c_f[3]);
/* free CLOVER */
QOP_CLOVER_free_gauge(&c_g);
for (i = 0; i < NELEMS(c_f); i++)
QOP_CLOVER_free_fermion(&c_f[i]);
QOP_CLOVER_fini(&clover_state);
/* Compute plaquette */
plaq = plaquette(U);
/* field norms */
for (i = 0; i < NELEMS(F); i++)
QDP_r_eq_norm2_D(&n[i], F[i], QDP_all);
/* Display the values */
printf0("plaquette = %g\n",
plaq / (QDP_volume() * QDP_Nc * NDIM * (NDIM - 1) / 2 ));
for (i = 0; i < NELEMS(F); i++)
printf0(" |f|^2 [%d] = %20.10e\n", i, (double)(n[i]));
/* Compute and display <f[1] f[2]> */
show_dot("1|D0", F[1], F[2]);
/* Compute and display <f[3] f[0]> */
show_dot("X1|0", F[3], F[0]);
QDP_destroy_S(state);
QDP_destroy_I(I_seed);
destroy_Mvector(U, NELEMS(U));
destroy_Mvector(C, NELEMS(C));
destroy_Dvector(F, NELEMS(F));
status = 0;
end:
/* shutdown QDP */
printf0("end\n");
QDP_finalize();
return status;
}
int
congrad_w(int niter, Real rsqmin, Real *final_rsq_ptr)
{
int i;
int iteration; /* counter for iterations */
double source_norm;
double rsqstop;
QLA_Real a, b;
double rsq,oldrsq,pkp; /* Sugar's a,b,resid**2,previous resid*2 */
/* pkp = cg_p.K.cg_p */
QLA_Real mkappa;
QLA_Real sum;
#ifdef CGTIME
double dtime;
#endif
#ifdef LU
mkappa = -kappa*kappa;
#else
mkappa = -kappa;
#endif
setup_cg();
for(i=0; i<4; i++) {
set_M_from_site(gaugelink[i], F_OFFSET(link[i]),EVENANDODD);
}
set_D_from_site(psi, F_OFFSET(psi),EVENANDODD);
set_D_from_site(chi, F_OFFSET(chi),EVENANDODD);
#ifdef PRESHIFT_LINKS
{
QDP_ColorMatrix *tcm;
tcm = QDP_create_M();
for(i=0; i<4; i++) {
QDP_M_eq_sM(tcm, gaugelink[i], QDP_neighbor[i], QDP_backward, QDP_all);
QDP_M_eq_Ma(gaugelink[i+4], tcm, QDP_all);
}
QDP_destroy_M(tcm);
}
#endif
#ifdef CGTIME
dtime = -dclock();
#endif
iteration=0;
start:
/* mp <- M_adjoint*M*psi
r,p <- chi - mp
rsq = |r|^2
source_norm = |chi|^2
*/
rsq = source_norm = 0.0;
#ifdef LU
QDP_D_eq_D(cgp, psi, QDP_even);
dslash_special_qdp(tt1, cgp, 1, QDP_odd, temp1);
dslash_special_qdp(ttt, tt1, 1, QDP_even, temp2);
QDP_D_eq_r_times_D_plus_D(ttt, &mkappa, ttt, cgp, QDP_even);
dslash_special_qdp(tt2, ttt, -1, QDP_odd, temp3);
dslash_special_qdp(mp, tt2, -1, QDP_even, temp4);
QDP_D_eq_r_times_D_plus_D(mp, &mkappa, mp, ttt, QDP_even);
QDP_D_eq_D_minus_D(cgr, chi, mp, QDP_even);
QDP_D_eq_D(cgp, cgr, QDP_even);
QDP_r_eq_norm2_D(&sum, chi, QDP_even);
source_norm = sum;
QDP_r_eq_norm2_D(&sum, cgr, QDP_even);
rsq = sum;
#else
QDP_D_eq_D(cgp, psi, QDP_even);
dslash_special_qdp(ttt, cgp, 1, QDP_all, temp1);
QDP_D_eq_r_times_D_plus_D(ttt, &mkappa, ttt, cgp, QDP_all);
dslash_special_qdp(mp, ttt, -1, QDP_all, temp1);
QDP_D_eq_r_times_D_plus_D(mp, &mkappa, mp, ttt, QDP_all);
QDP_D_eq_D_minus_D(cgr, chi, mp, QDP_all);
QDP_D_eq_D(cgp, cgr, QDP_all);
QDP_r_eq_norm2_D(&sum, chi, QDP_all);
source_norm = sum;
QDP_r_eq_norm2_D(&sum, cgr, QDP_all);
rsq = sum;
#endif
iteration++ ; /* iteration counts number of multiplications
by M_adjoint*M */
total_iters++;
/**if(this_node==0)printf("congrad2: source_norm = %e\n",source_norm);
if(this_node==0)printf("congrad2: iter %d, rsq %e, pkp %e, a %e\n",
iteration,(double)rsq,(double)pkp,(double)a );**/
rsqstop = rsqmin * source_norm;
if( rsq <= rsqstop ){
*final_rsq_ptr= (Real)rsq;
return (iteration);
}
/* main loop - do until convergence or time to restart */
/*
oldrsq <- rsq
mp <- M_adjoint*M*p
pkp <- p.M_adjoint*M.p
a <- rsq/pkp
psi <- psi + a*p
r <- r - a*mp
rsq <- |r|^2
b <- rsq/oldrsq
p <- r + b*p
*/
do {
oldrsq = rsq;
#ifdef LU
dslash_special_qdp(tt1, cgp, 1, QDP_odd, temp1);
dslash_special_qdp(ttt, tt1, 1, QDP_even, temp2);
QDP_D_eq_r_times_D_plus_D(ttt, &mkappa, ttt, cgp, QDP_even);
dslash_special_qdp(tt2, ttt, -1, QDP_odd, temp3);
dslash_special_qdp(mp, tt2, -1, QDP_even, temp4);
QDP_D_eq_r_times_D_plus_D(mp, &mkappa, mp, ttt, QDP_even);
QDP_r_eq_re_D_dot_D(&sum, cgp, mp, QDP_even);
pkp = sum;
#else
dslash_special_qdp(ttt, cgp, 1, QDP_all, temp1);
QDP_D_eq_r_times_D_plus_D(ttt, &mkappa, ttt, cgp, QDP_all);
dslash_special_qdp(mp, ttt, -1, QDP_all, temp1);
QDP_D_eq_r_times_D_plus_D(mp, &mkappa, mp, ttt, QDP_all);
QDP_r_eq_re_D_dot_D(&sum, cgp, mp, QDP_all);
pkp = sum;
#endif
iteration++;
total_iters++;
a = rsq / pkp;
QDP_D_peq_r_times_D(psi, &a, cgp, MYSUBSET);
QDP_D_meq_r_times_D(cgr, &a, mp, MYSUBSET);
QDP_r_eq_norm2_D(&sum, cgr, MYSUBSET);
rsq = sum;
/**if(this_node==0)printf("congrad2: iter %d, rsq %e, pkp %e, a %e\n",
iteration,(double)rsq,(double)pkp,(double)a );**/
if( rsq <= rsqstop ){
*final_rsq_ptr= (Real)rsq;
#ifdef CGTIME
dtime += dclock();
if(this_node==0)
printf("CONGRAD2: time = %.2e size_r= %.2e iters= %d MF = %.1f\n",
dtime,rsq,iteration,
(double)6480*iteration*even_sites_on_node/(dtime*1e6));
//(double)5616*iteration*even_sites_on_node/(dtime*1e6));
#endif
set_site_from_D(F_OFFSET(psi), psi,EVENANDODD);
return (iteration);
}
b = rsq / oldrsq;
QDP_D_eq_r_times_D_plus_D(cgp, &b, cgp, cgr, MYSUBSET);
} while( iteration%niter != 0);
set_site_from_D(F_OFFSET(psi), psi,EVENANDODD);
if( iteration < 3*niter ) goto start;
*final_rsq_ptr= (Real)rsq;
return(iteration);
}
