static void prune_el_on_cl
( mclMatrix* el_to_cl /* must be conforming */
, mclMatrix* el_on_cl /* this one will be pruned */
, double pct
, int max
)
{ dim i
; for (i=0;i<N_COLS(el_on_cl);i++)
{ mclv* elclvec = el_on_cl->cols+i
; long clid = el_to_cl->cols[i].ivps[0].idx
; double sum = 0.0
; int n_others = 0
; dim k = 0
; mcxbool selfok = FALSE
; mclvSort(elclvec, mclpValRevCmp)
; while (k++ < elclvec->n_ivps && sum < pct && n_others < max)
{ long y = elclvec->ivps[k-1].idx
; if (y == clid)
selfok = TRUE
; sum += elclvec->ivps[k-1].val
; n_others++
; }
mclvResize(elclvec, k-1) /* careful recentchange */
; mclvSort(elclvec, mclpIdxCmp)
; if (!selfok)
mclvInsertIdx(elclvec, clid, 0.01)
; }
}
void mclvSortAscVal
( mclVector* vec
)
{ mclvSort(vec, mclpValCmp)
; }
void mclvSortDescVal
( mclVector* vec
)
{ mclvSort(vec, mclpValRevCmp)
; }
void mclvSortUniq
( mclVector* vec
)
{ mclvSort(vec, mclpIdxCmp)
; mclvUniqIdx(vec, mclpMergeLeft)
; }
/* current dst content is thrown away if fltbinary not used */
mclv* mclvFromPAR
( mclv* dst
, mclpAR* par
, mcxbits warnbits
, void (*ivpmerge)(void* ivp1, const void* ivp2)
, double (*fltbinary)(pval val1, pval val2)
)
{ mcxbool warn_re = warnbits & MCLV_WARN_REPEAT_ENTRIES
; mcxbool warn_rv = warnbits & MCLV_WARN_REPEAT_VECTORS
; mclp* ivps = par->ivps
; dim n_ivps = par->n_ivps
; mcxbits sortbits = par->sorted
; dim n_old = dst ? dst->n_ivps : 0
; const char* me = "mclvFromPAR"
; dim n_re = 0, n_rv = 0
; if (!dst)
dst = mclvInit(NULL)
; if (n_ivps)
{ if (dst->n_ivps && fltbinary)
{ mclVector* tmpvec = mclvNew(ivps, n_ivps)
; if (!(sortbits & MCLPAR_SORTED))
mclvSort(tmpvec, NULL)
; if (!(sortbits & MCLPAR_UNIQUE))
n_re = mclvUniqIdx(tmpvec, ivpmerge)
; n_rv += tmpvec->n_ivps
; n_rv += dst->n_ivps
; mclvBinary(dst, tmpvec, dst, fltbinary)
; n_rv -= dst->n_ivps
; mclvFree(&tmpvec)
; }
else
{ if (dst->ivps == ivps)
mcxErr(me, "DANGER dst->ivps == ivps (dst vid %d)", (int) dst->vid)
; mclvRenew(dst, ivps, n_ivps)
; if (!(sortbits & MCLPAR_SORTED))
mclvSort(dst, NULL)
; if (!(sortbits & MCLPAR_UNIQUE))
n_re += mclvUniqIdx(dst, ivpmerge)
; }
}
if (warn_re && n_re)
mcxErr
( me
, "<%ld> found <%ld> repeated entries within %svector"
, (long) dst->vid
, (long) n_re
, n_rv ? "repeated " : ""
)
; if (warn_rv && n_rv)
mcxErr
( me
, "<%ld> new vector has <%ld> overlap with previous amalgam"
, (long) dst->vid
, (long) n_rv
)
; if (warnbits && n_re + n_rv)
mcxErr
( me
, "<%ld> vector went from <%ld> to <%ld> entries"
, (long) dst->vid
, (long) n_old
, (long) dst->n_ivps
)
; return dst
; }
static dim clm_clm_prune
( mclx* mx
, mclx* cl
, dim prune_sz
, mclx** cl_adjustedpp
, dim* n_sink
, dim* n_source
)
{ dim d, n_adjusted = 0
; mclx* cl_adj = mclxCopy(cl)
; mclv* cid_affected = mclvClone(cl->dom_cols)
; const char* me = "clmAssimilate"
; double bar_affected = 1.5
; mclx *el_to_cl = NULL
; mclx *el_on_cl = NULL
; mclx *cl_on_cl = NULL
; mclx *cl_on_el = NULL
; *n_sink = 0
; *n_source = 0
; mclvMakeConstant(cid_affected, 1.0)
; mclxColumnsRealign(cl_adj, mclvSizeCmp)
; *cl_adjustedpp = NULL
; clmCastActors
(&mx, &cl_adj, &el_to_cl, &el_on_cl, &cl_on_cl, &cl_on_el, 0.95)
; mclxFree(&cl_on_el)
; for (d=0;d<N_COLS(cl_on_cl);d++)
{ mclv* clthis = cl_adj->cols+d
; mclv* cllist = cl_on_cl->cols+d
; mclp* pself = mclvGetIvp(cllist, clthis->vid, NULL)
; double self_val = -1.0
; if (pself)
self_val = pself->val
, pself->val *= 1.001 /* to push it up in case of equal weights */
;if(0)fprintf(stderr, "test size %d\n", (int) clthis->n_ivps)
; if (prune_sz && clthis->n_ivps > prune_sz)
continue
; while (1)
{ mclv* clthat
; dim e
; if (cllist->n_ivps < 2)
break
; mclvSort(cllist, mclpValRevCmp)
/* now get biggest mass provided that cluster
* ranks higher (has at least as many entries)
*
* fixme/todo: we probably have a slight order
* dependency for some fringe cases. If provable
* then either solve or document it.
*/
; for (e=0;e<cllist->n_ivps;e++)
if (cllist->ivps[e].idx >= clthis->vid)
break
/* found none or itself */
; if (e == cllist->n_ivps || cllist->ivps[e].idx == clthis->vid)
break
; if /* Should Not Happen */
(!(clthat
= mclxGetVector(cl_adj, cllist->ivps[e].idx, RETURN_ON_FAIL, NULL)
) )
break
/* works for special case prune_sz == 0 */
/* if (clthat->n_ivps + clthis->n_ivps > prune_sz) */
/* ^iced. inconsistent behaviour as k grows. */
; { mcxLog
( MCX_LOG_LIST
, me
, "source %ld|%lu|%.3f absorbed by %ld|%lu|%.3f"
, clthis->vid, (ulong) clthis->n_ivps, self_val
, clthat->vid, (ulong) clthat->n_ivps, cllist->ivps[0].val
)
; n_adjusted += clthis->n_ivps
; (*n_sink)++
/* note: we could from our precomputed cl_on_cl
* obtain that A is absorbed in B, B is absorbed in C.
* below we see that A will be merged with B,
* and the result will then be merged with C.
* This depends on the fact that cl_adj is ordered
* on increasing cluster size.
*/
; mcldMerge(cl_adj->cols+d, clthat, clthat)
; mclvResize(cl_adj->cols+d, 0)
; mclvInsertIdx(cid_affected, clthat->vid, 2.0)
; }
break
; }
mclvSort(cllist, mclpIdxCmp)
; }
mclxFree(&cl_on_cl)
; mclxFree(&el_on_cl)
; mclxFree(&el_to_cl)
; mclxMakeCharacteristic(cl)
; mclvUnary(cid_affected, fltxGT, &bar_affected)
; *n_source = cid_affected->n_ivps
; mclvFree(&cid_affected)
; mclxColumnsRealign(cl_adj, mclvSizeRevCmp)
; if (!n_adjusted)
{ mclxFree(&cl_adj)
; return 0
; }
mclxUnary(cl_adj, fltxCopy, NULL)
; mclxMakeCharacteristic(cl_adj)
; *cl_adjustedpp = cl_adj
; return n_adjusted
; }
