#include "BSprivate.h"
/*@ BSbackward1 - Backward triangular matrix multiplication on a
single vector
Input Parameters:
. A - The sparse matrix
. x - The input vector
. comm - The communication structure for A
. procinfo - the usual processor information
Output Parameters:
. b - on exit contains A*x
Returns:
void
Notes:
We assume that A has no i-nodes or cliques
@*/
void BSbackward1(BSpar_mat *A, FLOAT *x, FLOAT *b,
BScomm *comm, BSprocinfo *procinfo)
{
BMcomp_msg *from_msg, *to_msg;
BMphase *to_phase, *from_phase;
BMmsg *msg;
int i, j, k;
int cl_ind, in_ind;
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;
if(!A->icc_storage) {
/* nonsymmetric version done in BSforward() */
return;
}
color2clique = A->color2clique;
clique2inode = A->clique2inode;
inodes = A->inodes->list;
/* REMEMBER, the to and from phase are switched here */
from_msg = comm->to_msg;
to_msg = comm->from_msg;
/* post for all messages */
BMinit_comp_msg(from_msg,procinfo); CHKERR(0);
/* REMEMBER, the diagonal has already been taken care */
/* now do this phase by phase */
for (i=color2clique->length-2;i>=0;i--) {
/* first send my messages */
/* this will involve computing partial sums */
to_phase = BMget_phase(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);
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 = 0.0;
for (k=0;k<size;k++) t += x[row[k]]*nz[k];
msg_buf[count] = t;
}
count++;
}
BMsendf_msg(msg,procinfo); CHKERR(0);
}
CHKERR(0);
}
/* do some local work */
for (i=color2clique->length-2;i>=0;i--) {
for (cl_ind=color2clique->numbers[i];
cl_ind<color2clique->numbers[i+1];cl_ind++) {
if (procinfo->my_id == clique2inode->proc[cl_ind]) {
/* only do the strictly upper triangular part */
/* we ASSUME the diagonal is all 1's */
/* which is taken care prior to this */
/* now, 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 = 0.0;
for (k=0;k<size;k++) t += x[row[k]]*nz[k];
b[ind] += t;
}
}
}
}
/* receive my messages and update my rhs */
for (i=color2clique->length-2;i>=0;i--) {
from_phase = BMget_phase(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);
msg_len = BMget_msg_size(msg); CHKERR(0);
count = 0;
for (j=0;j<msg_len;j++) b[data_ptr[j]] += msg_buf[j];
BMfree_msg(msg); CHKERR(0);
}
CHKERR(0);
}
/* wait for all of the sent messages to finish */
BMfinish_comp_msg(to_msg,procinfo); CHKERR(0);
MLOG_flop((2*A->local_nnz));
}
syntax highlighted by Code2HTML, v. 0.9.1