=================================================================== RCS file: /home/cvs/OpenXM_contrib2/asir2000/engine/nd.c,v retrieving revision 1.94 retrieving revision 1.116 diff -u -p -r1.94 -r1.116 --- OpenXM_contrib2/asir2000/engine/nd.c 2004/03/15 07:30:44 1.94 +++ OpenXM_contrib2/asir2000/engine/nd.c 2004/12/01 12:36:17 1.116 @@ -1,4 +1,4 @@ -/* $OpenXM: OpenXM_contrib2/asir2000/engine/nd.c,v 1.93 2004/03/13 07:48:30 noro Exp $ */ +/* $OpenXM: OpenXM_contrib2/asir2000/engine/nd.c,v 1.115 2004/11/18 08:29:11 noro Exp $ */ #include "nd.h" @@ -8,7 +8,9 @@ NM _nm_free_list; ND _nd_free_list; ND_pairs _ndp_free_list; +#if 0 static int ndv_alloc; +#endif #if 1 static int nd_f4_nsp=0x7fffffff; #else @@ -29,13 +31,19 @@ static NDV *nd_ps; static NDV *nd_ps_trace; static RHist *nd_psh; static int nd_psn,nd_pslen; - static RHist *nd_red; - +static int *nd_work_vector; +static int **nd_matrix; +static int nd_matrix_len; +static struct weight_or_block *nd_worb; +static int nd_worb_len; static int nd_found,nd_create,nd_notfirst; static int nmv_adv; static int nd_demand; +UINT *nd_det_compute_bound(NDV **dm,int n,int j); +void nd_det_reconstruct(NDV **dm,int n,int j,NDV d); + void nd_free_private_storage() { _nm_free_list = 0; @@ -422,6 +430,84 @@ int ndl_block_compare(UINT *d1,UINT *d2) return 0; } +int ndl_matrix_compare(UINT *d1,UINT *d2) +{ + int i,j,s; + int *v; + + for ( j = 0; j < nd_nvar; j++ ) + nd_work_vector[j] = GET_EXP(d1,j)-GET_EXP(d2,j); + for ( i = 0; i < nd_matrix_len; i++ ) { + v = nd_matrix[i]; + for ( j = 0, s = 0; j < nd_nvar; j++ ) + s += v[j]*nd_work_vector[j]; + if ( s > 0 ) return 1; + else if ( s < 0 ) return -1; + } + return 0; +} + +int ndl_composite_compare(UINT *d1,UINT *d2) +{ + int i,j,s,start,end,len,o; + int *v; + struct sparse_weight *sw; + + for ( j = 0; j < nd_nvar; j++ ) + nd_work_vector[j] = GET_EXP(d1,j)-GET_EXP(d2,j); + for ( i = 0; i < nd_worb_len; i++ ) { + len = nd_worb[i].length; + switch ( nd_worb[i].type ) { + case IS_DENSE_WEIGHT: + v = nd_worb[i].body.dense_weight; + for ( j = 0, s = 0; j < len; j++ ) + s += v[j]*nd_work_vector[j]; + if ( s > 0 ) return 1; + else if ( s < 0 ) return -1; + break; + case IS_SPARSE_WEIGHT: + sw = nd_worb[i].body.sparse_weight; + for ( j = 0, s = 0; j < len; j++ ) + s += sw[j].value*nd_work_vector[sw[j].pos]; + if ( s > 0 ) return 1; + else if ( s < 0 ) return -1; + break; + case IS_BLOCK: + o = nd_worb[i].body.block.order; + start = nd_worb[i].body.block.start; + switch ( o ) { + case 0: + end = start+len; + for ( j = start, s = 0; j < end; j++ ) + s += MUL_WEIGHT(nd_work_vector[j],j); + if ( s > 0 ) return 1; + else if ( s < 0 ) return -1; + for ( j = end-1; j >= start; j-- ) + if ( nd_work_vector[j] < 0 ) return 1; + else if ( nd_work_vector[j] > 0 ) return -1; + break; + case 1: + end = start+len; + for ( j = start, s = 0; j < end; j++ ) + s += MUL_WEIGHT(nd_work_vector[j],j); + if ( s > 0 ) return 1; + else if ( s < 0 ) return -1; + for ( j = start; j < end; j++ ) + if ( nd_work_vector[j] > 0 ) return 1; + else if ( nd_work_vector[j] < 0 ) return -1; + break; + case 2: + for ( j = start; j < end; j++ ) + if ( nd_work_vector[j] > 0 ) return 1; + else if ( nd_work_vector[j] < 0 ) return -1; + break; + } + break; + } + } + return 0; +} + /* TDH -> WW -> TD-> RL */ int ndl_ww_lex_compare(UINT *d1,UINT *d2) @@ -641,13 +727,11 @@ int ndl_disjoint(UINT *d1,UINT *d2) #endif } -int ndl_check_bound2(int index,UINT *d2) +int ndl_check_bound(UINT *d1,UINT *d2) { UINT u2; - UINT *d1; int i,j,ind,k; - d1 = nd_bound[index]; ind = 0; #if USE_UNROLL switch ( nd_bpe ) { @@ -736,6 +820,11 @@ int ndl_check_bound2(int index,UINT *d2) #endif } +int ndl_check_bound2(int index,UINT *d2) +{ + return ndl_check_bound(nd_bound[index],d2); +} + INLINE int ndl_hash_value(UINT *d) { int i; @@ -873,6 +962,79 @@ ND nd_add(int mod,ND p1,ND p2) } } +/* XXX on opteron, the inlined manipulation of destructive additon of + * two NM seems to make gcc optimizer get confused, so the part is + * done in a function. + */ + +int nm_destructive_add_q(NM *m1,NM *m2,NM *mr0,NM *mr) +{ + NM s; + Q t; + int can; + + addq(CQ(*m1),CQ(*m2),&t); + s = *m1; *m1 = NEXT(*m1); + if ( t ) { + can = 1; NEXTNM2(*mr0,*mr,s); CQ(*mr) = (t); + } else { + can = 2; FREENM(s); + } + s = *m2; *m2 = NEXT(*m2); FREENM(s); + return can; +} + +ND nd_add_q(ND p1,ND p2) +{ + int n,c,can; + ND r; + NM m1,m2,mr0,mr,s; + Q t; + + if ( !p1 ) return p2; + else if ( !p2 ) return p1; + else { + can = 0; + for ( n = NV(p1), m1 = BDY(p1), m2 = BDY(p2), mr0 = 0; m1 && m2; ) { + c = DL_COMPARE(DL(m1),DL(m2)); + switch ( c ) { + case 0: +#if defined(__x86_64__) + can += nm_destructive_add_q(&m1,&m2,&mr0,&mr); +#else + addq(CQ(m1),CQ(m2),&t); + s = m1; m1 = NEXT(m1); + if ( t ) { + can++; NEXTNM2(mr0,mr,s); CQ(mr) = (t); + } else { + can += 2; FREENM(s); + } + s = m2; m2 = NEXT(m2); FREENM(s); +#endif + break; + case 1: + s = m1; m1 = NEXT(m1); NEXTNM2(mr0,mr,s); + break; + case -1: + s = m2; m2 = NEXT(m2); NEXTNM2(mr0,mr,s); + break; + } + } + if ( !mr0 ) + if ( m1 ) mr0 = m1; + else if ( m2 ) mr0 = m2; + else return 0; + else if ( m1 ) NEXT(mr) = m1; + else if ( m2 ) NEXT(mr) = m2; + else NEXT(mr) = 0; + BDY(p1) = mr0; + SG(p1) = MAX(SG(p1),SG(p2)); + LEN(p1) = LEN(p1)+LEN(p2)-can; + FREEND(p2); + return p1; + } +} + ND nd_add_sf(ND p1,ND p2) { int n,c,can; @@ -1131,7 +1293,7 @@ again: d = ndvtond(0,r); stat = nd_nf(0,d,nd_ps,0,0,&nf); if ( !stat ) { - nd_reconstruct(0,0,0); + nd_reconstruct(0,0); goto again; } else if ( nf ) return 0; if ( DP_Print ) { printf("."); fflush(stdout); } @@ -1441,7 +1603,7 @@ again: stat = nd_sp(m,0,l,&h); if ( !stat ) { NEXT(l) = d; d = l; - d = nd_reconstruct(m,0,d); + d = nd_reconstruct(0,d); goto again; } #if USE_GEOBUCKET @@ -1451,7 +1613,7 @@ again: #endif if ( !stat ) { NEXT(l) = d; d = l; - d = nd_reconstruct(m,0,d); + d = nd_reconstruct(0,d); goto again; } else if ( nf ) { if ( checkonly ) return 0; @@ -1551,7 +1713,7 @@ again: stat = nd_sp(m,0,l,&h); if ( !stat ) { NEXT(l) = d; d = l; - d = nd_reconstruct(m,1,d); + d = nd_reconstruct(1,d); goto again; } #if USE_GEOBUCKET @@ -1561,7 +1723,7 @@ again: #endif if ( !stat ) { NEXT(l) = d; d = l; - d = nd_reconstruct(m,1,d); + d = nd_reconstruct(1,d); goto again; } else if ( nf ) { if ( nd_demand ) { @@ -1572,7 +1734,7 @@ again: if ( !nfq ) { if ( !nd_sp(0,1,l,&h) || !nd_nf(0,h,nd_ps_trace,!Top,0,&nfq) ) { NEXT(l) = d; d = l; - d = nd_reconstruct(m,1,d); + d = nd_reconstruct(1,d); goto again; } } @@ -1636,7 +1798,7 @@ NODE ndv_reduceall(int m,NODE f) g = nd_separate_head(g,&head); stat = nd_nf(m,g,nd_ps,1,&dn,&nf); if ( !stat ) - nd_reconstruct(m,0,0); + nd_reconstruct(0,0); else { if ( DP_Print ) { printf("."); fflush(stdout); } if ( !m ) { @@ -2032,7 +2194,9 @@ void nd_gr(LIST f,LIST v,int m,int f4,struct order_spe if ( !m && Demand ) nd_demand = 1; else nd_demand = 0; +#if 0 ndv_alloc = 0; +#endif get_vars((Obj)f,&fv); pltovl(v,&vv); for ( nvar = 0, tv = vv; tv; tv = NEXT(tv), nvar++ ); switch ( ord->id ) { @@ -2070,7 +2234,9 @@ void nd_gr(LIST f,LIST v,int m,int f4,struct order_spe } if ( r0 ) NEXT(r) = 0; MKLIST(*rp,r0); +#if 0 fprintf(asir_out,"ndv_alloc=%d\n",ndv_alloc); +#endif } void nd_gr_trace(LIST f,LIST v,int trace,int homo,struct order_spec *ord,LIST *rp) @@ -2419,7 +2585,7 @@ void removecont_array(Q *c,int n) { struct oVECT v; Q d0,d1,a,u,u1,gcd; - int i; + int i,j; N qn,rn,gn; Q *q,*r; @@ -2456,6 +2622,7 @@ void nd_mul_c(int mod,ND p,int mul) int c,c1; if ( !p ) return; + if ( mul == 1 ) return; if ( mod == -1 ) for ( m = BDY(p); m; m = NEXT(m) ) CM(m) = _mulsf(CM(m),mul); @@ -2471,6 +2638,7 @@ void nd_mul_c_q(ND p,Q mul) Q c; if ( !p ) return; + if ( UNIQ(mul) ) return; for ( m = BDY(p); m; m = NEXT(m) ) { mulq(CQ(m),mul,&c); CQ(m) = c; } @@ -2548,17 +2716,46 @@ UINT *ndv_compute_bound(NDV p) return t; } +UINT *nd_compute_bound(ND p) +{ + UINT *d1,*d2,*t; + UINT u; + int i,j,k,l,len,ind; + NM m; + + if ( !p ) + return 0; + d1 = (UINT *)ALLOCA(nd_wpd*sizeof(UINT)); + d2 = (UINT *)ALLOCA(nd_wpd*sizeof(UINT)); + len = LEN(p); + m = BDY(p); ndl_copy(DL(m),d1); m = NEXT(m); + for ( m = NEXT(m); m; m = NEXT(m) ) { + ndl_lcm(DL(m),d1,d2); + t = d1; d1 = d2; d2 = t; + } + l = nd_nvar+31; + t = (UINT *)MALLOC_ATOMIC(l*sizeof(UINT)); + for ( i = nd_exporigin, ind = 0; i < nd_wpd; i++ ) { + u = d1[i]; + k = (nd_epw-1)*nd_bpe; + for ( j = 0; j < nd_epw; j++, k -= nd_bpe, ind++ ) + t[ind] = (u>>k)&nd_mask0; + } + for ( ; ind < l; ind++ ) t[ind] = 0; + return t; +} + int nd_get_exporigin(struct order_spec *ord) { switch ( ord->id ) { - case 0: + case 0: case 2: return 1; case 1: /* block order */ /* d[0]:weight d[1]:w0,...,d[nd_exporigin-1]:w(n-1) */ return ord->ord.block.length+1; - case 2: - error("nd_get_exporigin : matrix order is not supported yet."); + case 3: + error("nd_get_exporigin : composite order is not supported yet."); } } @@ -2603,9 +2800,10 @@ void nd_setup_parameters(int nvar,int max) { nmv_adv = ROUND_FOR_ALIGN(sizeof(struct oNMV)+(nd_wpd-1)*sizeof(UINT)); nd_epos = nd_create_epos(nd_ord); nd_blockmask = nd_create_blockmask(nd_ord); + nd_work_vector = (int *)REALLOC(nd_work_vector,nd_nvar*sizeof(int)); } -ND_pairs nd_reconstruct(int mod,int trace,ND_pairs d) +ND_pairs nd_reconstruct(int trace,ND_pairs d) { int i,obpe,oadv,h; static NM prev_nm_free_list; @@ -3012,6 +3210,54 @@ ND ndv_mul_nm(int mod,NM m0,NDV p) } } +ND nd_quo(int mod,PGeoBucket bucket,NDV d) +{ + NM mq0,mq; + NMV tm; + Q q; + int i,nv,sg,c,c1,c2,hindex; + ND p,t,r; + N tnm; + + if ( !p ) return 0; + else { + nv = NV(d); + mq0 = 0; + tm = (NMV)ALLOCA(nmv_adv); + while ( 1 ) { + hindex = mod?head_pbucket(mod,bucket):head_pbucket_q(bucket); + if ( hindex < 0 ) break; + p = bucket->body[hindex]; + NEXTNM(mq0,mq); + ndl_sub(HDL(p),HDL(d),DL(mq)); + ndl_copy(DL(mq),DL(tm)); + if ( mod ) { + c1 = invm(HCM(d),mod); c2 = HCM(p); + DMAR(c1,c2,0,mod,c); CM(mq) = c; + CM(tm) = mod-c; + } else { + divsn(NM(HCQ(p)),NM(HCQ(d)),&tnm); + NTOQ(tnm,SGN(HCQ(p))*SGN(HCQ(d)),CQ(mq)); + chsgnq(CQ(mq),&CQ(tm)); + } + t = ndv_mul_nmv_trunc(mod,tm,d,HDL(d)); + bucket->body[hindex] = nd_remove_head(p); + t = nd_remove_head(t); + add_pbucket(mod,bucket,t); + } + if ( !mq0 ) + r = 0; + else { + NEXT(mq) = 0; + for ( i = 0, mq = mq0; mq; mq = NEXT(mq), i++ ); + MKND(nv,mq0,i,r); + /* XXX */ + SG(r) = HTD(r); + } + return r; + } +} + void ndv_realloc(NDV p,int obpe,int oadv,EPOS oepos) { NMV m,mr,mr0,t; @@ -3161,7 +3407,7 @@ ND ptond(VL vl,VL dvl,P p) w = (DCP *)ALLOCA(k*sizeof(DCP)); for ( dc = DC(p), j = 0; j < k; dc = NEXT(dc), j++ ) w[j] = dc; for ( i = 0, tvl = dvl, v = VR(p); - vl && tvl->v != v; tvl = NEXT(tvl), i++ ); + tvl && tvl->v != v; tvl = NEXT(tvl), i++ ); if ( !tvl ) { for ( j = k-1, s = 0, MKV(v,x); j >= 0; j-- ) { t = ptond(vl,dvl,COEF(w[j])); @@ -3234,7 +3480,9 @@ NDV ndtondv(int mod,ND p) m0 = m = (NMV)GC_malloc_atomic_ignore_off_page(len*nmv_adv); else m0 = m = MALLOC(len*nmv_adv); +#if 0 ndv_alloc += nmv_adv*len; +#endif for ( t = BDY(p), i = 0; t; t = NEXT(t), i++, NMV_ADV(m) ) { ndl_copy(DL(t),DL(m)); CQ(m) = CQ(t); @@ -3354,14 +3602,29 @@ void nd_init_ord(struct order_spec *ord) } break; case 1: + /* block order */ /* XXX */ nd_dcomp = -1; nd_isrlex = 0; ndl_compare_function = ndl_block_compare; break; case 2: - error("nd_init_ord : matrix order is not supported yet."); + /* matrix order */ + /* XXX */ + nd_dcomp = -1; + nd_isrlex = 0; + nd_matrix_len = ord->ord.matrix.row; + nd_matrix = ord->ord.matrix.matrix; + ndl_compare_function = ndl_matrix_compare; break; + case 3: + /* composite order */ + nd_dcomp = -1; + nd_isrlex = 0; + nd_worb_len = ord->ord.composite.length; + nd_worb = ord->ord.composite.w_or_b; + ndl_compare_function = ndl_composite_compare; + break; } nd_ord = ord; } @@ -3372,7 +3635,8 @@ BlockMask nd_create_blockmask(struct order_spec *ord) UINT *t; BlockMask bm; - if ( !ord->id ) + /* we only create mask table for block order */ + if ( ord->id != 1 ) return 0; n = ord->ord.block.length; bm = (BlockMask)MALLOC(sizeof(struct oBlockMask)); @@ -3430,7 +3694,14 @@ EPOS nd_create_epos(struct order_spec *ord) } break; case 2: - error("nd_create_epos : matrix order is not supported yet."); + /* matrix order */ + case 3: + /* composite order */ + for ( i = 0; i < nd_nvar; i++ ) { + epos[i].i = nd_exporigin + i/nd_epw; + epos[i].s = (nd_epw-(i%nd_epw)-1)*nd_bpe; + } + break; } return epos; } @@ -3446,6 +3717,10 @@ void nd_nf_p(P f,LIST g,LIST v,int m,struct order_spec int stat,nvar,max,e; union oNDC dn; + if ( !f ) { + *rp = 0; + return; + } pltovl(v,&vv); for ( nvar = 0, tv = vv; tv; tv = NEXT(tv), nvar++ ); @@ -3482,7 +3757,7 @@ void nd_nf_p(P f,LIST g,LIST v,int m,struct order_spec stat = nd_nf(m,nd,nd_ps,1,0,&nf); if ( !stat ) { nd_psn++; - nd_reconstruct(m,0,0); + nd_reconstruct(0,0); nd_psn--; } else break; @@ -3506,47 +3781,22 @@ int nd_to_vect(int mod,UINT *s0,int n,ND d,UINT *r) return i; } -int ndv_to_vect(int mod,UINT *s0,int n,NDV d,UINT *r) +int nd_to_vect_q(UINT *s0,int n,ND d,Q *r) { - NMV m; + NM m; UINT *t,*s; - int i,j,len; + int i; for ( i = 0; i < n; i++ ) r[i] = 0; - m = BDY(d); - len = LEN(d); - for ( i = j = 0, s = s0; j < len; j++, NMV_ADV(m)) { + for ( i = 0, s = s0, m = BDY(d); m; m = NEXT(m) ) { t = DL(m); for ( ; !ndl_equal(t,s); s += nd_wpd, i++ ); - r[i] = CM(m); + r[i] = CQ(m); } for ( i = 0; !r[i]; i++ ); return i; } -int nm_ind_pair_to_vect(int mod,UINT *s0,int n,NM_ind_pair pair,UINT *r) -{ - NM m; - NMV mr; - UINT *d,*t,*s; - NDV p; - int i,j,len; - - m = pair->mul; - d = DL(m); - p = nd_ps[pair->index]; - t = (UINT *)ALLOCA(nd_wpd*sizeof(UINT)); - for ( i = 0; i < n; i++ ) r[i] = 0; - len = LEN(p); - for ( i = j = 0, s = s0, mr = BDY(p); j < len; j++, NMV_ADV(mr) ) { - ndl_add(d,DL(mr),t); - for ( ; !ndl_equal(t,s); s += nd_wpd, i++ ); - r[i] = CM(mr); - } - for ( i = 0; !r[i]; i++ ); - return i; -} - IndArray nm_ind_pair_to_vect_compress(int mod,UINT *s0,int n,NM_ind_pair pair) { NM m; @@ -3596,6 +3846,63 @@ IndArray nm_ind_pair_to_vect_compress(int mod,UINT *s0 } +int ndv_reduce_vect_q(Q *svect,int col,IndArray *imat,NM_ind_pair *rp0,int nred) +{ + int i,j,k,len,pos,prev; + Q cs,mcs,c1,c2,cr,gcd,t; + IndArray ivect; + unsigned char *ivc; + unsigned short *ivs; + unsigned int *ivi; + NDV redv; + NMV mr; + NODE rp; + int maxrs; + + maxrs = 0; + for ( i = 0; i < nred; i++ ) { + ivect = imat[i]; + k = ivect->head; + if ( svect[k] ) { + maxrs = MAX(maxrs,rp0[i]->sugar); + redv = nd_ps[rp0[i]->index]; + len = LEN(redv); mr = BDY(redv); + igcd_cofactor(svect[k],CQ(mr),&gcd,&cs,&cr); + chsgnq(cs,&mcs); + if ( !UNIQ(cr) ) { + for ( j = 0; j < col; j++ ) { + mulq(svect[j],cr,&c1); svect[j] = c1; + } + } + svect[k] = 0; prev = k; + switch ( ivect->width ) { + case 1: + ivc = ivect->index.c; + for ( j = 1, NMV_ADV(mr); j < len; j++, NMV_ADV(mr) ) { + pos = prev+ivc[j]; prev = pos; + mulq(CQ(mr),mcs,&c2); addq(svect[pos],c2,&t); svect[pos] = t; + } + break; + case 2: + ivs = ivect->index.s; + for ( j = 1, NMV_ADV(mr); j < len; j++, NMV_ADV(mr) ) { + pos = prev+ivs[j]; prev = pos; + mulq(CQ(mr),mcs,&c2); addq(svect[pos],c2,&t); svect[pos] = t; + } + break; + case 4: + ivi = ivect->index.i; + for ( j = 1, NMV_ADV(mr); j < len; j++, NMV_ADV(mr) ) { + pos = prev+ivi[j]; prev = pos; + mulq(CQ(mr),mcs,&c2); addq(svect[pos],c2,&t); svect[pos] = t; + } + break; + } + } + } + return maxrs; +} + int ndv_reduce_vect(int m,UINT *svect,int col,IndArray *imat,NM_ind_pair *rp0,int nred) { int i,j,k,len,pos,prev; @@ -3719,7 +4026,9 @@ NDV vect_to_ndv(UINT *vect,int spcol,int col,int *rhea if ( !len ) return 0; else { mr0 = (NMV)GC_malloc_atomic_ignore_off_page(nmv_adv*len); +#if 0 ndv_alloc += nmv_adv*len; +#endif mr = mr0; p = s0vect; for ( j = k = 0; j < col; j++, p += nd_wpd ) @@ -3733,6 +4042,36 @@ NDV vect_to_ndv(UINT *vect,int spcol,int col,int *rhea } } +NDV vect_to_ndv_q(Q *vect,int spcol,int col,int *rhead,UINT *s0vect) +{ + int j,k,len; + UINT *p; + Q c; + NDV r; + NMV mr0,mr; + + for ( j = 0, len = 0; j < spcol; j++ ) if ( vect[j] ) len++; + if ( !len ) return 0; + else { + mr0 = (NMV)GC_malloc(nmv_adv*len); +#if 0 + ndv_alloc += nmv_adv*len; +#endif + mr = mr0; + p = s0vect; + for ( j = k = 0; j < col; j++, p += nd_wpd ) + if ( !rhead[j] ) { + if ( c = vect[k++] ) { + if ( DN(c) ) + error("afo"); + ndl_copy(p,DL(mr)); CQ(mr) = c; NMV_ADV(mr); + } + } + MKNDV(nd_nvar,mr0,len,r); + return r; + } +} + int nd_sp_f4(int m,ND_pairs l,PGeoBucket bucket) { ND_pairs t; @@ -3812,9 +4151,9 @@ NODE nd_f4(int m) PGeoBucket bucket; struct oEGT eg0,eg1,eg_f4; - if ( !m ) - error("nd_f4 : not implemented"); +#if 0 ndv_alloc = 0; +#endif g = 0; d = 0; for ( i = 0; i < nd_psn; i++ ) { d = update_pairs(d,g,i); @@ -3829,7 +4168,7 @@ NODE nd_f4(int m) if ( !stat ) { for ( t = l; NEXT(t); t = NEXT(t) ); NEXT(t) = d; d = l; - d = nd_reconstruct(m,0,d); + d = nd_reconstruct(0,d); continue; } if ( bucket->m < 0 ) continue; @@ -3837,7 +4176,7 @@ NODE nd_f4(int m) if ( !col ) { for ( t = l; NEXT(t); t = NEXT(t) ); NEXT(t) = d; d = l; - d = nd_reconstruct(m,0,d); + d = nd_reconstruct(0,d); continue; } get_eg(&eg1); init_eg(&eg_f4); add_eg(&eg_f4,&eg0,&eg1); @@ -3858,30 +4197,23 @@ NODE nd_f4(int m) } } for ( r = g; r; r = NEXT(r) ) BDY(r) = (pointer)nd_ps[(int)BDY(r)]; +#if 0 fprintf(asir_out,"ndv_alloc=%d\n",ndv_alloc); +#endif return g; } NODE nd_f4_red(int m,ND_pairs sp0,UINT *s0vect,int col,NODE rp0) { IndArray *imat; - int nsp,nred,spcol,sprow,a; + int nsp,nred,i; int *rhead; - int i,j,k,l,rank; - NODE rp,r0,r; + NODE r0,rp; ND_pairs sp; - ND spol; - int **spmat; - UINT *svect,*v; - int *colstat; - struct oEGT eg0,eg1,eg2,eg_f4,eg_f4_1,eg_f4_2; NM_ind_pair *rvect; - int maxrs; - int *spsugar; - get_eg(&eg0); for ( sp = sp0, nsp = 0; sp; sp = NEXT(sp), nsp++ ); - nred = length(rp0); spcol = col-nred; + nred = length(rp0); imat = (IndArray *)ALLOCA(nred*sizeof(IndArray)); rhead = (int *)ALLOCA(col*sizeof(int)); for ( i = 0; i < col; i++ ) rhead[i] = 0; @@ -3893,7 +4225,30 @@ NODE nd_f4_red(int m,ND_pairs sp0,UINT *s0vect,int col imat[i] = nm_ind_pair_to_vect_compress(m,s0vect,col,rvect[i]); rhead[imat[i]->head] = 1; } + if ( m ) + r0 = nd_f4_red_main(m,sp0,nsp,s0vect,col,rvect,rhead,imat,nred); + else + r0 = nd_f4_red_q_main(sp0,nsp,s0vect,col,rvect,rhead,imat,nred); + return r0; +} +NODE nd_f4_red_main(int m,ND_pairs sp0,int nsp,UINT *s0vect,int col, + NM_ind_pair *rvect,int *rhead,IndArray *imat,int nred) +{ + int spcol,sprow,a; + int i,j,k,l,rank; + NODE r0,r; + ND_pairs sp; + ND spol; + int **spmat; + UINT *svect,*v; + int *colstat; + struct oEGT eg0,eg1,eg2,eg_f4,eg_f4_1,eg_f4_2; + int maxrs; + int *spsugar; + + spcol = col-nred; + get_eg(&eg0); /* elimination (1st step) */ spmat = (int **)ALLOCA(nsp*sizeof(UINT *)); svect = (UINT *)ALLOCA(col*sizeof(UINT)); @@ -3937,6 +4292,7 @@ NODE nd_f4_red(int m,ND_pairs sp0,UINT *s0vect,int col SG((NDV)BDY(r)) = spsugar[i]; GC_free(spmat[i]); } + if ( r0 ) NEXT(r) = 0; for ( ; i < sprow; i++ ) GC_free(spmat[i]); get_eg(&eg2); init_eg(&eg_f4_2); add_eg(&eg_f4_2,&eg1,&eg2); init_eg(&eg_f4); add_eg(&eg_f4,&eg0,&eg2); @@ -3949,6 +4305,74 @@ NODE nd_f4_red(int m,ND_pairs sp0,UINT *s0vect,int col return r0; } +NODE nd_f4_red_q_main(ND_pairs sp0,int nsp,UINT *s0vect,int col, + NM_ind_pair *rvect,int *rhead,IndArray *imat,int nred) +{ + int spcol,sprow,a; + int i,j,k,l,rank; + NODE r0,r; + ND_pairs sp; + ND spol; + Q **spmat; + Q *svect,*v; + int *colstat; + struct oEGT eg0,eg1,eg2,eg_f4,eg_f4_1,eg_f4_2; + int maxrs; + int *spsugar; + + spcol = col-nred; + get_eg(&eg0); + /* elimination (1st step) */ + spmat = (Q **)ALLOCA(nsp*sizeof(Q *)); + svect = (Q *)ALLOCA(col*sizeof(Q)); + spsugar = (int *)ALLOCA(nsp*sizeof(Q)); + for ( a = sprow = 0, sp = sp0; a < nsp; a++, sp = NEXT(sp) ) { + nd_sp(0,0,sp,&spol); + if ( !spol ) continue; + nd_to_vect_q(s0vect,col,spol,svect); + maxrs = ndv_reduce_vect_q(svect,col,imat,rvect,nred); + for ( i = 0; i < col; i++ ) if ( svect[i] ) break; + if ( i < col ) { + spmat[sprow] = v = (Q *)MALLOC(spcol*sizeof(Q)); + for ( j = k = 0; j < col; j++ ) + if ( !rhead[j] ) v[k++] = svect[j]; + spsugar[sprow] = MAX(maxrs,SG(spol)); + sprow++; + } +/* nd_free(spol); */ + } + get_eg(&eg1); init_eg(&eg_f4_1); add_eg(&eg_f4_1,&eg0,&eg1); + if ( DP_Print ) { + fprintf(asir_out,"elim1=%fsec,",eg_f4_1.exectime+eg_f4_1.gctime); + fflush(asir_out); + } + /* free index arrays */ +/* for ( i = 0; i < nred; i++ ) GC_free(imat[i]->index.c); */ + + /* elimination (2nd step) */ + colstat = (int *)ALLOCA(spcol*sizeof(int)); + rank = nd_gauss_elim_q(spmat,spsugar,sprow,spcol,colstat); + r0 = 0; + for ( i = 0; i < rank; i++ ) { + NEXTNODE(r0,r); BDY(r) = + (pointer)vect_to_ndv_q(spmat[i],spcol,col,rhead,s0vect); + SG((NDV)BDY(r)) = spsugar[i]; +/* GC_free(spmat[i]); */ + } + if ( r0 ) NEXT(r) = 0; + +/* for ( ; i < sprow; i++ ) GC_free(spmat[i]); */ + get_eg(&eg2); init_eg(&eg_f4_2); add_eg(&eg_f4_2,&eg1,&eg2); + init_eg(&eg_f4); add_eg(&eg_f4,&eg0,&eg2); + if ( DP_Print ) { + fprintf(asir_out,"elim2=%fsec\n",eg_f4_2.exectime+eg_f4_2.gctime); + fprintf(asir_out,"nsp=%d,nred=%d,spmat=(%d,%d),rank=%d ", + nsp,nred,sprow,spcol,rank); + fprintf(asir_out,"%fsec\n",eg_f4.exectime+eg_f4.gctime); + } + return r0; +} + FILE *nd_write,*nd_read; void nd_send_int(int a) { @@ -4012,7 +4436,9 @@ NDV nd_recv_ndv() if ( !len ) return 0; else { m0 = m = (NMV)GC_malloc_atomic_ignore_off_page(nmv_adv*len); +#if 0 ndv_alloc += len*nmv_adv; +#endif for ( i = 0; i < len; i++, NMV_ADV(m) ) { CM(m) = nd_recv_int(); nd_recv_intarray(DL(m),nd_wpd); @@ -4203,6 +4629,58 @@ void nd_exec_f4_red_dist() fflush(nd_write); } +int nd_gauss_elim_q(Q **mat0,int *sugar,int row,int col,int *colstat) +{ + int mod,i,j,t,c,rank,rank0,inv; + int *ci,*ri; + Q dn; + MAT m,nm; + int **wmat; + + /* XXX */ + mod = 99999989; + wmat = (int **)ALLOCA(row*sizeof(int *)); + for ( i = 0; i < row; i++ ) { + wmat[i] = (int *)ALLOCA(col*sizeof(int)); + for ( j = 0; j < col; j++ ) { + if ( mat0[i][j] ) { + t = rem(NM(mat0[i][j]),mod); + if ( SGN(mat0[i][j]) < 0 ) t = mod-t; + wmat[i][j] = t; + } else + wmat[i][j] = 0; + } + } + rank0 = nd_gauss_elim_mod(wmat,sugar,row,col,mod,colstat); + NEWMAT(m); m->row = row; m->col = col; m->body = (pointer **)mat0; + rank = generic_gauss_elim(m,&nm,&dn,&ri,&ci); + if ( rank != rank0 ) + error("afo"); + for ( i = 0; i < row; i++ ) + for ( j = 0; j < col; j++ ) + mat0[i][j] = 0; + c = col-rank; + for ( i = 0; i < rank; i++ ) { + mat0[i][ri[i]] = dn; + for ( j = 0; j < c; j++ ) + mat0[i][ci[j]] = (Q)BDY(nm)[i][j]; + } + inv = invm(rem(NM(dn),mod),mod); + if ( SGN(dn) < 0 ) inv = mod-inv; + for ( i = 0; i < row; i++ ) + for ( j = 0; j < col; j++ ) { + if ( mat0[i][j] ) { + t = rem(NM(mat0[i][j]),mod); + if ( SGN(mat0[i][j]) < 0 ) t = mod-t; + } else + t = 0; + c = dmar(t,inv,0,mod); + if ( wmat[i][j] != c ) + error("afo"); + } + return rank; +} + int nd_gauss_elim_mod(int **mat0,int *sugar,int row,int col,int md,int *colstat) { int i,j,k,l,inv,a,rank,s; @@ -4408,4 +4886,213 @@ NDV ndv_load(int index) MKNDV(nv,m0,len,d); SG(d) = sugar; return d; +} + +void nd_det(int mod,MAT f,P *rp) +{ + VL fv,tv; + int n,i,j,max,e,nvar,sgn,k0,l0,len0,len,k,l,a; + pointer **m; + Q mone; + NDV **dm; + NDV *t,*mi,*mj; + NDV d,s,mij,mjj; + ND u; + NMV nmv; + UINT *bound; + PGeoBucket bucket; + struct order_spec *ord; + + create_order_spec(0,0,&ord); + nd_init_ord(ord); + get_vars((Obj)f,&fv); + if ( f->row != f->col ) + error("nd_det : non-square matrix"); + n = f->row; + for ( nvar = 0, tv = fv; tv; tv = NEXT(tv), nvar++ ); + m = f->body; + for ( i = 0, max = 0; i < n; i++ ) + for ( j = 0; j < n; j++ ) + for ( tv = fv; tv; tv = NEXT(tv) ) { + e = getdeg(tv->v,(P)m[i][j]); + max = MAX(e,max); + } + nd_setup_parameters(nvar,max); + dm = (NDV **)almat_pointer(n,n); + for ( i = 0, max = 0; i < n; i++ ) + for ( j = 0; j < n; j++ ) { + dm[i][j] = ptondv(CO,fv,m[i][j]); + if ( mod ) ndv_mod(mod,dm[i][j]); + if ( dm[i][j] && !LEN(dm[i][j]) ) dm[i][j] = 0; + } + d = ptondv(CO,fv,(P)ONE); + if ( mod ) ndv_mod(mod,d); + chsgnq(ONE,&mone); + for ( j = 0, sgn = 1; j < n; j++ ) { + if ( DP_Print ) fprintf(stderr,"j=%d\n",j); + for ( i = j; i < n && !dm[i][j]; i++ ); + if ( i == n ) { + *rp = 0; + return; + } + k0 = i; l0 = j; len0 = LEN(dm[k0][l0]); + for ( k = j; k < n; k++ ) + for ( l = j; l < n; l++ ) + if ( dm[k][l] && LEN(dm[k][l]) < len0 ) { + k0 = k; l0 = l; len0 = LEN(dm[k][l]); + } + if ( k0 != j ) { + t = dm[j]; dm[j] = dm[k0]; dm[k0] = t; + sgn = -sgn; + } + if ( l0 != j ) { + for ( k = j; k < n; k++ ) { + s = dm[k][j]; dm[k][j] = dm[k][l0]; dm[k][l0] = s; + } + sgn = -sgn; + } + bound = nd_det_compute_bound(dm,n,j); + if ( ndl_check_bound(bound,bound) ) + nd_det_reconstruct(dm,n,j,d); + + for ( i = j+1, mj = dm[j], mjj = mj[j]; i < n; i++ ) { +/* if ( DP_Print ) fprintf(stderr," i=%d\n ",i); */ + mi = dm[i]; mij = mi[j]; + if ( mod ) + ndv_mul_c(mod,mij,mod-1); + else + ndv_mul_c_q(mij,mone); + for ( k = j+1; k < n; k++ ) { +/* if ( DP_Print ) fprintf(stderr,"k=%d ",k); */ + bucket = create_pbucket(); + if ( mi[k] ) { + nmv = BDY(mjj); len = LEN(mjj); + for ( a = 0; a < len; a++, NMV_ADV(nmv) ) { + u = ndv_mul_nmv_trunc(mod,nmv,mi[k],DL(BDY(d))); + add_pbucket(mod,bucket,u); + } + } + if ( mj[k] && mij ) { + nmv = BDY(mij); len = LEN(mij); + for ( a = 0; a < len; a++, NMV_ADV(nmv) ) { + u = ndv_mul_nmv_trunc(mod,nmv,mj[k],DL(BDY(d))); + add_pbucket(mod,bucket,u); + } + } + u = nd_quo(mod,bucket,d); + mi[k] = ndtondv(mod,u); + } +/* if ( DP_Print ) fprintf(stderr,"\n",k); */ + } + d = mjj; + } + if ( sgn < 0 ) + if ( mod ) + ndv_mul_c(mod,d,mod-1); + else + ndv_mul_c_q(d,mone); + *rp = ndvtop(mod,CO,fv,d); +} + +ND ndv_mul_nmv_trunc(int mod,NMV m0,NDV p,UINT *d) +{ + NM mr,mr0; + NM tnm; + NMV m; + UINT *d0,*dt,*dm; + int c,n,td,i,c1,c2,len; + Q q; + ND r; + + if ( !p ) return 0; + else { + n = NV(p); m = BDY(p); len = LEN(p); + d0 = DL(m0); + td = TD(d); + mr0 = 0; + NEWNM(tnm); + if ( mod ) { + c = CM(m0); + for ( i = 0; i < len; i++, NMV_ADV(m) ) { + ndl_add(DL(m),d0,DL(tnm)); + if ( ndl_reducible(DL(tnm),d) ) { + NEXTNM(mr0,mr); + c1 = CM(m); DMAR(c1,c,0,mod,c2); CM(mr) = c2; + ndl_copy(DL(tnm),DL(mr)); + } + } + } else { + q = CQ(m0); + for ( i = 0; i < len; i++, NMV_ADV(m) ) { + ndl_add(DL(m),d0,DL(tnm)); + if ( ndl_reducible(DL(tnm),d) ) { + NEXTNM(mr0,mr); + mulq(CQ(m),q,&CQ(mr)); + ndl_copy(DL(tnm),DL(mr)); + } + } + } + if ( !mr0 ) + return 0; + else { + NEXT(mr) = 0; + for ( len = 0, mr = mr0; mr; mr = NEXT(mr), len++ ); + MKND(NV(p),mr0,len,r); + SG(r) = SG(p) + TD(d0); + return r; + } + } +} + +void nd_det_reconstruct(NDV **dm,int n,int j,NDV d) +{ + int i,obpe,oadv,h,k,l; + static NM prev_nm_free_list; + EPOS oepos; + + obpe = nd_bpe; + oadv = nmv_adv; + oepos = nd_epos; + if ( obpe < 2 ) nd_bpe = 2; + else if ( obpe < 3 ) nd_bpe = 3; + else if ( obpe < 4 ) nd_bpe = 4; + else if ( obpe < 5 ) nd_bpe = 5; + else if ( obpe < 6 ) nd_bpe = 6; + else if ( obpe < 8 ) nd_bpe = 8; + else if ( obpe < 10 ) nd_bpe = 10; + else if ( obpe < 16 ) nd_bpe = 16; + else if ( obpe < 32 ) nd_bpe = 32; + else error("nd_det_reconstruct : exponent too large"); + + nd_setup_parameters(nd_nvar,0); + prev_nm_free_list = _nm_free_list; + _nm_free_list = 0; + for ( k = j; k < n; k++ ) + for (l = j; l < n; l++ ) + ndv_realloc(dm[k][l],obpe,oadv,oepos); + ndv_realloc(d,obpe,oadv,oepos); + prev_nm_free_list = 0; +#if 0 + GC_gcollect(); +#endif +} + +UINT *nd_det_compute_bound(NDV **dm,int n,int j) +{ + UINT *d0,*d1,*d,*t,*r; + int k,l; + + d0 = (UINT *)ALLOCA(nd_wpd*sizeof(UINT)); + d1 = (UINT *)ALLOCA(nd_wpd*sizeof(UINT)); + for ( k = 0; k < nd_wpd; k++ ) d0[k] = 0; + for ( k = j; k < n; k++ ) + for ( l = j; l < n; l++ ) + if ( dm[k][l] ) { + d = ndv_compute_bound(dm[k][l]); + ndl_lcm(d,d0,d1); + t = d1; d1 = d0; d0 = t; + } + r = (UINT *)ALLOCA(nd_wpd*sizeof(UINT)); + for ( k = 0; k < nd_wpd; k++ ) r[k] = d0[k]; + return r; }