int merge_ordered_sets (const int **labels,const float **vals,
const int *sizes,int k,
int **labels_out,float **vals_out) {
int i,j;
int n_out = ivec_sum (sizes, k);
int *all_labels = ivec_new (n_out);
float *all_vals = fvec_new (n_out);
/* Maxheap:
* * maxheap label = index of table in 0..k-1
* * maxheap val = - (label from labels table)
*
* If maxheap val does not fit in a float (if label>2**24), it
* triggers an assertion. Time to implement a maxheap with int
* values...
*/
fbinheap_t *mh = fbinheap_new(k);
/* current index on table k */
int indices[k];
for ( i = 0 ; i < k ; i++) {
if (sizes[i] == 0)
continue;
indices[i] = 0;
int label = labels[i][0];
float mh_val = -label;
assert ((int)(-mh_val) == label || !"lost precision in int->float conversion");
fbinheap_add (mh, i, mh_val);
}
int all_i = 0;
while (mh->k>0) {
/* smallest available label */
i = mh->label[1]; /* index of table */
j = (int)(-mh->val[1]); /* label */
/* I don't dare compiling with -DNDEBUG */
/* assert(j==labels[i][indices[i]]); */
all_labels[all_i] = j;
all_vals[all_i] = vals[i][indices[i]];
all_i++;
/* remove handled label */
fbinheap_pop (mh);
indices[i]++;
if (indices[i] < sizes[i]) { /* push next label from this table */
int label = labels[i][indices[i]];
float mh_val = -label;
assert ((int)(-mh_val) == label || !"lost precision in int->float conversion");
fbinheap_add (mh, i, mh_val);
}
}
fbinheap_delete (mh);
assert (all_i == n_out);
*labels_out = all_labels;
*vals_out = all_vals;
return n_out;
}
static PetscErrorCode set_pairwise(gs_id *gs)
{
PetscInt i, j;
PetscInt p_mask_size;
PetscInt *p_mask, *sh_proc_mask, *tmp_proc_mask;
PetscInt *ngh_buf, *buf2;
PetscInt offset;
PetscInt *msg_list, *msg_size, **msg_nodes, nprs;
PetscInt *pairwise_elm_list, len_pair_list=0;
PetscInt *iptr, t1, i_start, nel, *elms;
PetscInt ct;
PetscErrorCode ierr;
PetscFunctionBegin;
/* to make life easier */
nel = gs->nel;
elms = gs->elms;
ngh_buf = gs->ngh_buf;
sh_proc_mask = gs->pw_nghs;
/* need a few temp masks */
p_mask_size = len_bit_mask(num_nodes);
p_mask = (PetscInt*) malloc(p_mask_size);
tmp_proc_mask = (PetscInt*) malloc(p_mask_size);
/* set mask to my my_id's bit mask */
ierr = set_bit_mask(p_mask,p_mask_size,my_id);CHKERRQ(ierr);
p_mask_size /= sizeof(PetscInt);
len_pair_list=gs->len_pw_list;
gs->pw_elm_list=pairwise_elm_list=(PetscInt*)malloc((len_pair_list+1)*sizeof(PetscInt));
/* how many processors (nghs) do we have to exchange with? */
nprs=gs->num_pairs=ct_bits((char *)sh_proc_mask,p_mask_size*sizeof(PetscInt));
/* allocate space for gs_gop() info */
gs->pair_list = msg_list = (PetscInt *) malloc(sizeof(PetscInt)*nprs);
gs->msg_sizes = msg_size = (PetscInt *) malloc(sizeof(PetscInt)*nprs);
gs->node_list = msg_nodes = (PetscInt **) malloc(sizeof(PetscInt*)*(nprs+1));
/* init msg_size list */
ierr = ivec_zero(msg_size,nprs);CHKERRQ(ierr);
/* expand from bit mask list to int list */
ierr = bm_to_proc((char *)sh_proc_mask,p_mask_size*sizeof(PetscInt),msg_list);CHKERRQ(ierr);
/* keep list of elements being handled pairwise */
for (i=j=0;i<nel;i++)
{
if (elms[i] & TOP_BIT)
{elms[i] ^= TOP_BIT; pairwise_elm_list[j++] = i;}
}
pairwise_elm_list[j] = -1;
gs->msg_ids_out = (MPI_Request *) malloc(sizeof(MPI_Request)*(nprs+1));
gs->msg_ids_out[nprs] = MPI_REQUEST_NULL;
gs->msg_ids_in = (MPI_Request *) malloc(sizeof(MPI_Request)*(nprs+1));
gs->msg_ids_in[nprs] = MPI_REQUEST_NULL;
gs->pw_vals = (PetscScalar *) malloc(sizeof(PetscScalar)*len_pair_list*vec_sz);
/* find who goes to each processor */
for (i_start=i=0;i<nprs;i++)
{
/* processor i's mask */
ierr = set_bit_mask(p_mask,p_mask_size*sizeof(PetscInt),msg_list[i]);CHKERRQ(ierr);
/* det # going to processor i */
for (ct=j=0;j<len_pair_list;j++)
{
buf2 = ngh_buf+(pairwise_elm_list[j]*p_mask_size);
ierr = ivec_and3(tmp_proc_mask,p_mask,buf2,p_mask_size);CHKERRQ(ierr);
if (ct_bits((char *)tmp_proc_mask,p_mask_size*sizeof(PetscInt)))
{ct++;}
}
msg_size[i] = ct;
i_start = PetscMax(i_start,ct);
/*space to hold nodes in message to first neighbor */
msg_nodes[i] = iptr = (PetscInt*) malloc(sizeof(PetscInt)*(ct+1));
for (j=0;j<len_pair_list;j++)
{
buf2 = ngh_buf+(pairwise_elm_list[j]*p_mask_size);
ierr = ivec_and3(tmp_proc_mask,p_mask,buf2,p_mask_size);CHKERRQ(ierr);
if (ct_bits((char *)tmp_proc_mask,p_mask_size*sizeof(PetscInt)))
{*iptr++ = j;}
}
*iptr = -1;
}
msg_nodes[nprs] = NULL;
j=gs->loc_node_pairs=i_start;
t1 = GL_MAX;
ierr = giop(&i_start,&offset,1,&t1);CHKERRQ(ierr);
gs->max_node_pairs = i_start;
i_start=j;
t1 = GL_MIN;
ierr = giop(&i_start,&offset,1,&t1);CHKERRQ(ierr);
gs->min_node_pairs = i_start;
i_start=j;
t1 = GL_ADD;
ierr = giop(&i_start,&offset,1,&t1);CHKERRQ(ierr);
gs->avg_node_pairs = i_start/num_nodes + 1;
i_start=nprs;
t1 = GL_MAX;
giop(&i_start,&offset,1,&t1);
gs->max_pairs = i_start;
/* remap pairwise in tail of gsi_via_bit_mask() */
gs->msg_total = ivec_sum(gs->msg_sizes,nprs);
gs->out = (PetscScalar *) malloc(sizeof(PetscScalar)*gs->msg_total*vec_sz);
gs->in = (PetscScalar *) malloc(sizeof(PetscScalar)*gs->msg_total*vec_sz);
/* reset malloc pool */
free((void*)p_mask);
free((void*)tmp_proc_mask);
PetscFunctionReturn(0);
}
