void fix_tableidtopic(int d, int t) {
int val;
assert(ddS.Tdt);
ddS.Tdt[d][t]++;
val = atomic_incr(ddS.TDt[t]);
if ( val==1 )
atomic_incr(ddS.TDTnz);
atomic_incr(ddS.TDT);
}
void fix_tableidword(int w, int t) {
int val;
val = atomic_incr(ddS.Twt[w][t]);
atomic_incr(ddS.TWt[t]);
val = atomic_incr(ddS.TwT[w]);
if ( val==1 ) {
atomic_incr( ddS.TWTnz);
}
atomic_incr(ddS.TWT);
}
/*
* add affects of document to stats
*
* when using multis, reset ddM.Mi[] to be totals
* for training words only, ignore test words in hold
*/
int add_doc(int d, enum GibbsType fix) {
int i, t, w, nd=0;
int mi = 0;
if ( ddP.bdk!=NULL )
mi = ddM.MI[d];
for (i=ddD.NdTcum[d]; i<ddD.NdTcum[d+1]; i++) {
if ( fix!=GibbsHold || pctl_hold(i) ) {
/*
* these words are for perp. calcs
*/
nd++;
}
if ( fix!=GibbsHold || !pctl_hold(i) ) {
t = Z_t(ddS.z[i]);
/*
* these words are for training
*/
ddS.Ndt[d][t]++;
ddS.NdT[d]++;
if ( (ddP.bdk==NULL) || Z_issetr(ddS.z[i]) ) {
if ( ddP.phi==NULL ) {
int val;
w = ddD.w[i];
atomic_incr(ddS.NWt[t]);
val = atomic_incr(ddS.Nwt[w][t]);
if ( ddP.PYbeta && val==1 )
fix_tableidword(w,t);
}
}
}
if ( (ddP.bdk!=NULL) && M_multi(i) ) mi++;
}
if ( ddP.PYalpha ) {
/* initialise ddS.Tdt[d][*] */
/*
* adjust table count stats based on Ndt[d]
*/
for (t=0; t<ddN.T; t++) {
if ( ddS.Ndt[d][t]>0 ) {
int val;
ddS.Tdt[d][t] = 1;
val = atomic_incr(ddS.TDt[t]);
if ( val==1 )
atomic_incr(ddS.TDTnz);
atomic_incr(ddS.TDT);
} else
ddS.Tdt[d][t] = 0;
}
}
// yap_message("add_doc(%d): %d\n", d, nd);
return nd;
}
void *sampling_p(void *pargs)
{
int i;
float *p = fvec(ddN.T * 4);
D_MiSi_t dD;
D_pargs_p *par =(D_pargs_p *) pargs;
clock_t t1 = clock();
if ( PCTL_BURSTY() )
misi_init(&ddM,&dD);
/*
* sampling
*/
par->thislp = 0;
par->thisNd = 0;
while ( (i=atomic_incr(*par->doc)-1)<ddN.DT ) {
if ( PCTL_BURSTY() )
misi_build(&dD,i,0);
par->thislp += gibbs_lda(GibbsNone, i, ddD.N_dT[i], p, &dD);
par->thisNd += ddD.N_dT[i];
if ( par->dots>0 && i>0 && (i%par->dots==0) )
yap_message(".");
if ( PCTL_BURSTY() )
misi_unbuild(&dD,i,0);
}
free(p);
if ( PCTL_BURSTY() )
misi_free(&dD);
par->tot_time = (double)(clock() - t1) / CLOCKS_PER_SEC;
return NULL;
}
struct sockbase *getlistener (const char *addr) {
int refed = false;
struct str_rbe *entry;
struct sockbase *sb = 0;
mutex_lock (&xgb.lock);
if ((entry = str_rb_find (&xgb.inproc_listeners, addr, strlen (addr)))) {
sb = & (cont_of (entry, struct inproc_sock, lhentry))->base;
mutex_lock (&sb->lock);
if (!sb->flagset.epipe) {
refed = true;
atomic_incr (&sb->ref);
}
mutex_unlock (&sb->lock);
if (!refed)
sb = 0;
}
void unfix_tableidword(int w, int t) {
int val;
assert(ddS.Twt[w][t]);
val = atomic_decr(ddS.Twt[w][t]);
if ( val>UINT16_MAX-5 ) {
/*
* whoops
*/
atomic_incr(ddS.Twt[w][t]);
return;
}
val = atomic_decr(ddS.TwT[w]);
atomic_decr(ddS.TWt[t]);
atomic_decr(ddS.TWT);
if ( val==0 ) {
atomic_decr(ddS.TWTnz);
}
}
/*
* got a new topic, so update all statistics;
*
*/
void update_topic(int i, int did, int wid, int t, int mi,
float dtip, D_MiSi_t *dD) {
int e = ddD.e[did];
int rd;
int rw=0;
/*
* fix up doc side
*/
rd = doc_side_ind(did, t);
assert(rd>=-1 && rd<=e+1);
if ( PCTL_BURSTY() )
// set wid negative if the word is bursty
wid = misi_incr(dD, i, mi, t, wid, dtip);
if ( rd>=0 ) {
/*
* increment affect of table indicator for topic PYP;
* when multicore, need to maintain consistency of constraints
*/
int n = ddS.n_dt[did][t];
if ( 0 && n==0 ) {
/* ripple to one, so go down chain */
fix_tableidtopic(did, t, rd);
ddS.N_dT[did]++;
ddS.n_dt[did][t] = 1;
} else {
/* increment up chain */
ddS.N_dT[did]++;
ddS.n_dt[did][t] = n+1;
fix_tableidtopic(did, t, rd);
}
} else {
ddS.N_dT[did]++;
ddS.n_dt[did][t]++;
}
/*
* fix up topicXword side
*/
if ( wid>=0 ) {
if ( ddP.phi )
rw = 0;
else
rw = word_side_ind(e, wid, t);
assert(rw>=0 && rw<=e+1);
if ( rw>0 ) {
/*
* subtract affect of table indicator for word PYP;
* when multicore, need to maintain consistency of constraints
*/
int zero = 0;
int incr = 0;
if ( atomic_incr_val(ddS.m_vte[wid][t][e],zero) ) {
if ( 1 ) {
atomic_incr(ddS.M_Vte[t][e]);
incr = 1;
} else {
atomic_decr(ddS.m_vte[wid][t][e]);
}
/* cancel increment and do afterwards */
} else {
atomic_incr(ddS.M_Vte[t][e]);
atomic_incr(ddS.m_vte[wid][t][e]);
incr = 1;
}
/*
* we have a new table for the word matrix
*/
#ifdef PHI_CACHE
{
int laste = fix_tableidword(e,wid,t,rw);
phi_sum_change(t,laste,i+1);
phi_norm_change(wid,t,laste);
}
#else
fix_tableidword(e,wid,t,rw);
#endif
if ( !incr ) {
atomic_incr(ddS.M_Vte[t][e]);
atomic_incr(ddS.m_vte[wid][t][e]);
}
} else {
/* simple case, order doesn't matter */
atomic_incr(ddS.M_Vte[t][e]);
atomic_incr(ddS.m_vte[wid][t][e]);
#ifdef PHI_CACHE
phi_norm_change(wid,t,e);
phi_sum_change(t,e,i+1);
#endif
}
}
#ifdef IND_STATS
if ( did < ddN.DT ) {
if ( rd>=0 && rd<=e )
atomic_incr(ddP.doc_ind_stats[t][e][rd]);
if ( wid>=0 && rw<=e )
atomic_incr(ddP.word_ind_stats[t][e][rw]);
}
#endif
}
