#include "BSprivate.h"
/*@ BSfor_solve1 - Forward triangular matrix solution on a
single vector
Input Parameters:
. A - The sparse matrix
. x - The rhs
. comm - The communication structure for A
. procinfo - the usual processor information
Output Parameters:
. x - on exit contains the solution vector
Returns:
void
Notes:
We assume that A has no i-nodes or cliques
@*/
void BSfor_solve1(BSpar_mat *A, FLOAT *x, BScomm *comm, BSprocinfo *procinfo)
{
BMphase *to_phase, *from_phase;
BMmsg *msg;
int i, j, k;
int cl_ind, in_ind, symmetric;
int count, size, ind;
int *row;
FLOAT *nz;
BScl_2_inode *clique2inode;
BSnumbering *color2clique;
BSinode *inodes;
int *data_ptr, msg_len;
FLOAT *msg_buf;
FLOAT t;
int *gnum, *iperm;
/* Is the symmetric data structure used? */
symmetric = A->icc_storage;
color2clique = A->color2clique;
clique2inode = A->clique2inode;
inodes = A->inodes->list;
gnum = A->global_row_num->numbers;
iperm = A->inv_perm->perm;
/* post for all messages */
BMinit_comp_msg(comm->from_msg,procinfo); CHKERR(0);
/* now do this phase by phase */
for (i=0;i<color2clique->length-1;i++) {
if(symmetric) {
/* find my portion of the solution using the cliques on the diagonal */
for (cl_ind=color2clique->numbers[i];
cl_ind<color2clique->numbers[i+1];cl_ind++) {
if (procinfo->my_id == clique2inode->proc[cl_ind]) {
/* first, multiply the clique */
/* the clique is stored, inverted, in the upper triangle */
ind = clique2inode->d_mats[cl_ind].local_ind;
x[ind] *= *(clique2inode->d_mats[cl_ind].matrix);
}
}
}
/* now send my messages */
to_phase = BMget_phase(comm->to_msg,i); CHKERR(0);
msg = NULL;
while ((msg = BMnext_msg(to_phase,msg)) != NULL) {
CHKERR(0);
msg_buf = (FLOAT *) BMget_msg_ptr(msg); CHKERR(0);
data_ptr = BMget_user(msg,&msg_len); CHKERR(0);
for (j=0;j<msg_len;j++) {
msg_buf[j] = x[data_ptr[j]];
}
BMsendf_msg(msg,procinfo); CHKERR(0);
}
CHKERR(0);
/* do some local work */
for (cl_ind=color2clique->numbers[i];
cl_ind<color2clique->numbers[i+1];cl_ind++) {
if (procinfo->my_id == clique2inode->proc[cl_ind]) {
/* multiply the inodes */
in_ind=clique2inode->inode_index[cl_ind];
size = inodes[in_ind].length;
if (size > 0) {
ind = clique2inode->d_mats[cl_ind].local_ind;
row = inodes[in_ind].row_num;
nz = inodes[in_ind].nz;
t = x[ind];
if(symmetric) {
for (k=0;k<size;k++) x[row[k]] -= t*nz[k];
} else {
for (k=0;k<size;k++) {
if (gnum[iperm[row[k]]] > inodes[in_ind].gcol_num)
x[row[k]] -= nz[k]*t;
}
}
}
}
}
/* receive my messages and do non-local work */
from_phase = BMget_phase(comm->from_msg,i); CHKERR(0);
while ((msg = BMrecv_msg(from_phase)) != NULL) {
CHKERR(0);
msg_buf = (FLOAT *) BMget_msg_ptr(msg); CHKERR(0);
data_ptr = BMget_user(msg,&msg_len); CHKERR(0);
count = 0;
for (cl_ind=data_ptr[0];cl_ind<=data_ptr[1];cl_ind++) {
in_ind=clique2inode->inode_index[cl_ind];
size = inodes[in_ind].length;
if (size > 0) {
row = inodes[in_ind].row_num;
nz = inodes[in_ind].nz;
t = msg_buf[count];
if(symmetric) {
for (k=0;k<size;k++) x[row[k]] -= t*nz[k];
} else {
for (k=0;k<size;k++) {
if (gnum[iperm[row[k]]] > inodes[in_ind].gcol_num)
x[row[k]] -= nz[k]*t;
}
}
}
count++;
}
BMfree_msg(msg); CHKERR(0);
}
CHKERR(0);
}
/* wait for all of the sent messages to finish */
BMfinish_comp_msg(comm->to_msg,procinfo); CHKERR(0);
MLOG_flop((2*A->local_nnz));
}
syntax highlighted by Code2HTML, v. 0.9.1