void dwf_dslash_5_plus_dag1(Vector *out,
Vector *in,
Float mass,
Dwf *dwf_lib_arg)
{
#pragma omp parallel default(shared)
{
// Initializations
//------------------------------------------------------------------
int idx;
IFloat *f_in;
IFloat *f_out;
int x;
int s;
const IFloat two_over_a5 = 2.0 * GJP.DwfA5Inv();
const IFloat neg_mass_two_over_a5 = -2.0 * mass * GJP.DwfA5Inv();
const int local_ls = GJP.SnodeSites();
const int s_nodes = GJP.Snodes();
const int s_node_coor = GJP.SnodeCoor();
const int vol_4d_cb = dwf_lib_arg->vol_4d / 2;
const int max_dex((local_ls-1)*vol_4d_cb);
const int ls_stride = 24 * vol_4d_cb;
IFloat *comm_buf = dwf_lib_arg->comm_buf;
// [1 + gamma_5] term (if dag=1 [1 - gamma_5] term)
//
// out[s] = [1 + gamma_5] in[s-1]
//------------------------------------------------------------------
f_in = (IFloat *) in;
f_out = (IFloat *) out;
//if(dag == 1){
f_in = f_in + 12;
f_out = f_out + 12;
//}
f_out = f_out + ls_stride;
#pragma omp for schedule(static)
for (idx=0;idx<max_dex;idx++)
{
cblas_daxpy(12,two_over_a5,f_in+24*idx,f_out+24*idx);
}
// [1 + gamma_5] for lower boundary term (if dag=1 [1 - gamma_5] term)
// If there's only one node along fifth direction, no communication
// is necessary; Otherwise data from adjacent node in minus direction
// will be needed.
// If the lower boundary is the s=0 term
// out[0] = - m_f * [1 + gamma_5] in[ls-1]
// else, out[s] = [1 + gamma_5] in[s-1]
//
//------------------------------------------------------------------
f_in = (IFloat *) in;
f_in = f_in + (local_ls-1)*ls_stride;
f_out = (IFloat *) out;
//if(dag == 1){
f_in = f_in + 12;
f_out = f_out + 12;
//}
#pragma omp for schedule(static)
for(x=0; x<vol_4d_cb; x++)
{
int shift(24*x);
IFloat *f_temp(f_in+shift);
if (s_nodes != 1 ) {
f_temp = comm_buf;
getMinusData(f_temp, f_in+shift, 12, 4);
}
if(s_node_coor == 0) {
cblas_daxpy(12,neg_mass_two_over_a5,f_temp,f_out+shift);
}
else {
cblas_daxpy(12,two_over_a5,f_temp,f_out+shift);
}
}
// [1 - gamma_5] term (if dag=1 [1 + gamma_5] term)
//
// out[s] = [1 - gamma_5] in[s+1]
//------------------------------------------------------------------
f_in = (IFloat *) in;
f_out = (IFloat *) out;
#if 0
if(dag == 0){
f_in = f_in + 12;
f_out = f_out + 12;
}
#endif
f_in = f_in + ls_stride;
#pragma omp for schedule(static)
for (idx=0;idx<max_dex;idx++)
{
const int shift(24*idx);
cblas_daxpy(12,two_over_a5,f_in+shift,f_out+shift);
}
// [1 - gamma_5] for upper boundary term (if dag=1 [1 + gamma_5] term)
// If there's only one node along fifth direction, no communication
// is necessary; Otherwise data from adjacent node in minus direction
// will be needed.
// If the upper boundary is the s=ls term
// out[ls-1] = - m_f * [1 - gamma_5] in[0]
// else out[s] = [1 - gamma_5] in[s+1]
//
//------------------------------------------------------------------
f_in = (IFloat *) in;
f_out = (IFloat *) out;
#if 0
if(dag == 0){
f_in = f_in + 12;
f_out = f_out + 12;
}
#endif
f_out = f_out + (local_ls-1)*ls_stride;
#pragma omp for schedule(static)
for(x=0; x<vol_4d_cb; x++){
const int shift(24*x);
IFloat *f_temp (f_in+shift);
if (s_nodes != 1 ) {
f_temp = comm_buf;
getPlusData(f_temp, f_in+shift, 12, 4);
}
if(s_node_coor == s_nodes - 1) {
cblas_daxpy(12,neg_mass_two_over_a5,f_temp,f_out+shift);
}
else {
cblas_daxpy(12,two_over_a5,f_temp,f_out+shift);
}
}
} // omp parallel
const int local_ls = GJP.SnodeSites();
const int vol_4d_cb = dwf_lib_arg->vol_4d / 2;
DiracOp::CGflops+=2*2*vol_4d_cb*local_ls*12;
}
void pSubChain::calc_acc_body()
{
// if(!children[1]) return;
#ifdef USE_MPI
if(sim->rank != rank) return;
if(children[0] && sim->rank != children[0]->rank)
{
children[0]->recv_acc();
}
if(children[1] && children[0] != children[1] && sim->rank != children[1]->rank)
{
children[1]->recv_acc();
}
#endif
// update_log << "--- " << last_joint->name << ": calc_acc_body" << endl;
int i, j;
// compute f_temp
static fVec da;
static fMat PK;
PK.resize(6, n_dof);
da.resize(n_const);
for(i=0; i<6; i++)
{
for(j=0; j<n_dof; j++)
PK(i, j) = P(i, joint_index[j]);
}
if(last_joint->n_dof > 0)
{
switch(last_joint->j_type)
{
case JROTATE:
case JSLIDE:
tau(0) = last_joint->tau;
break;
case JSPHERE:
tau(0) = last_joint->tau_n(0);
tau(1) = last_joint->tau_n(1);
tau(2) = last_joint->tau_n(2);
break;
case JFREE:
tau(0) = last_joint->tau_f(0);
tau(1) = last_joint->tau_f(1);
tau(2) = last_joint->tau_f(2);
tau(3) = last_joint->tau_n(0);
tau(4) = last_joint->tau_n(1);
tau(5) = last_joint->tau_n(2);
break;
default:
break;
}
da6.mul(PK, tau);
}
else
da6.zero();
// cerr << "da6(0) = " << tran(da6) << endl;
// + child_side - parent_side ?
da6 += children[0]->acc_temp[last_index[0]];
// cerr << "da6(1) = " << tran(da6) << endl;
if(children[1])
da6 -= children[1]->acc_temp[last_index[1]];
// cerr << "da6(2) = " << tran(da6) << endl;
// motion controlled joints
if(!last_joint->t_given)
{
switch(last_joint->j_type)
{
case JROTATE:
case JSLIDE:
da6(axis) -= last_joint->qdd;
// update_log << last_joint->name << ": qdd = " << last_joint->qdd << endl;
break;
case JSPHERE:
da6(3) -= last_joint->rel_ang_acc(0);
da6(4) -= last_joint->rel_ang_acc(1);
da6(5) -= last_joint->rel_ang_acc(2);
break;
case JFREE:
da6(0) -= last_joint->rel_lin_acc(0);
da6(1) -= last_joint->rel_lin_acc(1);
da6(2) -= last_joint->rel_lin_acc(2);
da6(3) -= last_joint->rel_ang_acc(0);
da6(4) -= last_joint->rel_ang_acc(1);
da6(5) -= last_joint->rel_ang_acc(2);
break;
}
}
static fVec f(6);
// cerr << "Gamma = " << Gamma << endl;
// cerr << "Gamma_inv = " << Gamma_inv << endl;
#if 0
// actually we could save some computation by
// selecting const rows first
for(i=0; i<n_const; i++)
da(i) = -da6(const_index[i]);
f_temp.mul(Gamma_inv, da);
// f_temp.lineq_posv(Gamma, da);
// compute acc at all outer joints
for(i=0; i<n_dof; i++)
f(joint_index[i]) = tau(i);
for(i=0; i<n_const; i++)
f(const_index[i]) = f_temp(i);
// cerr << "da = " << tran(da) << endl;
// cerr << "f_temp = " << tran(f_temp) << endl;
// cerr << "Gamma*f_temp - da = " << tran(Gamma*f_temp-da) << endl;
#else
#if 0
f.mul(W, da6);
for(i=0; i<n_dof; i++)
{
f(joint_index[i]) += tau(i);
}
#else
static fVec db(6), Wdb(6), IWRtau(6);
static fMat IWR;
IWR.resize(6, n_dof);
db.set(children[0]->acc_temp[last_index[0]]);
if(children[1])
db -= children[1]->acc_temp[last_index[1]];
Wdb.mul(W, db);
for(i=0; i<6; i++)
{
for(j=0; j<n_dof; j++)
{
IWR(i, j) = IW(joint_index[j], i);
}
}
IWRtau.mul(IWR, tau);
// cerr << "W = " << tran(W) << endl;
// update_log << "db = " << tran(db) << endl;
// cerr << "Wdb = " << tran(Wdb) << endl;
// cerr << "IWRtau = " << tran(IWRtau) << endl;
f.add(Wdb, IWRtau);
// update_log << "f = " << tran(f) << endl;
#ifdef PSIM_TEST
////// -> test
for(i=0; i<n_const; i++)
{
da(i) = -da6(const_index[i]);
f_temp(i) = f(const_index[i]);
}
// cerr << "Gamma*f_temp - da = " << tran(Gamma*f_temp-da) << endl;
total_gamma_error += (Gamma*f_temp-da) * (Gamma*f_temp-da);
////// <-
#endif
#endif
#endif
for(i=0; i<n_outer_joints; i++)
{
int org = outer_joints_origin[i];
int index = outer_joints_index[i];
int ilast = last_index[org];
acc_temp[i].mul(children[org]->Lambda[index][ilast], f);
if(org == 1)
{
acc_temp[i] *= -1.0;
}
acc_temp[i] += children[org]->acc_temp[index];
// update_log << "acc_temp[" << i << "] = " << tran(acc_temp[i]) << endl;
}
#ifdef USE_MPI
if(parent && sim->rank != parent->rank)
{
send_acc(parent->rank);
}
#endif
}