=================================================================== RCS file: /home/cvs/OpenXM_contrib2/asir2000/engine/nd.c,v retrieving revision 1.103 retrieving revision 1.114 diff -u -p -r1.103 -r1.114 --- OpenXM_contrib2/asir2000/engine/nd.c 2004/09/15 06:06:42 1.103 +++ OpenXM_contrib2/asir2000/engine/nd.c 2004/10/25 04:19:50 1.114 @@ -1,4 +1,4 @@ -/* $OpenXM: OpenXM_contrib2/asir2000/engine/nd.c,v 1.102 2004/09/15 01:43:33 noro Exp $ */ +/* $OpenXM: OpenXM_contrib2/asir2000/engine/nd.c,v 1.113 2004/10/06 12:09:19 noro Exp $ */ #include "nd.h" @@ -41,6 +41,9 @@ 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; @@ -724,13 +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 ) { @@ -819,6 +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; @@ -2579,7 +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; @@ -3202,21 +3208,24 @@ ND ndv_mul_nm(int mod,NM m0,NDV p) } } -ND nd_quo(int mod,ND p,NDV d) +ND nd_quo(int mod,PGeoBucket bucket,NDV d) { NM mq0,mq; NMV tm; Q q; - int i,nv,sg,c,c1,c2; - ND t,r; - + int i,nv,sg,c,c1,c2,hindex; + ND p,t,r; + N tnm; + if ( !p ) return 0; else { - nv = NV(p); - sg = SG(p); + nv = NV(d); mq0 = 0; tm = (NMV)ALLOCA(nmv_adv); - while ( p ) { + 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)); @@ -3225,17 +3234,24 @@ ND nd_quo(int mod,ND p,NDV d) DMAR(c1,c2,0,mod,c); CM(mq) = c; CM(tm) = mod-c; } else { - divq(HCQ(p),HCQ(d),&CQ(mq)); + 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)); - p = nd_add(mod,p,t); + bucket->body[hindex] = nd_remove_head(p); + t = nd_remove_head(t); + add_pbucket(mod,bucket,t); } - NEXT(mq) = 0; - for ( i = 0, mq = mq0; mq; mq = NEXT(mq), i++ ); - MKND(nv,mq0,i,r); - /* XXX */ - SG(r) = sg-SG(d); + 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; } } @@ -3759,47 +3775,22 @@ 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; @@ -3849,6 +3840,63 @@ 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; @@ -3988,6 +4036,36 @@ 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; @@ -4067,8 +4145,6 @@ 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 @@ -4124,23 +4200,14 @@ NODE nd_f4(int m) 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; @@ -4152,7 +4219,30 @@ 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)); @@ -4196,6 +4286,7 @@ 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); @@ -4208,6 +4299,74 @@ 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) { @@ -4464,6 +4623,58 @@ 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; @@ -4682,6 +4893,7 @@ void nd_det(int mod,MAT f,P *rp) NDV d,s,mij,mjj; ND u; NMV nmv; + UINT *bound; PGeoBucket bucket; struct order_spec *ord; @@ -4699,7 +4911,7 @@ void nd_det(int mod,MAT f,P *rp) e = getdeg(tv->v,(P)m[i][j]); max = MAX(e,max); } - nd_setup_parameters(nvar,1024); + 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++ ) { @@ -4711,6 +4923,7 @@ void nd_det(int mod,MAT f,P *rp) 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; @@ -4732,20 +4945,27 @@ void nd_det(int mod,MAT f,P *rp) } 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] ) + 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) ) { @@ -4753,10 +4973,10 @@ void nd_det(int mod,MAT f,P *rp) add_pbucket(mod,bucket,u); } } - u = normalize_pbucket(mod,bucket); - u = nd_quo(mod,u,d); + u = nd_quo(mod,bucket,d); mi[k] = ndtondv(mod,u); } +/* if ( DP_Print ) fprintf(stderr,"\n",k); */ } d = mjj; } @@ -4792,7 +5012,7 @@ ND ndv_mul_nmv_trunc(int mod,NMV m0,NDV p,UINT *d) if ( ndl_reducible(DL(tnm),d) ) { NEXTNM(mr0,mr); c1 = CM(m); DMAR(c1,c,0,mod,c2); CM(mr) = c2; - ndl_add(DL(m),d0,DL(mr)); + ndl_copy(DL(tnm),DL(mr)); } } } else { @@ -4802,7 +5022,7 @@ ND ndv_mul_nmv_trunc(int mod,NMV m0,NDV p,UINT *d) if ( ndl_reducible(DL(tnm),d) ) { NEXTNM(mr0,mr); mulq(CQ(m),q,&CQ(mr)); - ndl_add(DL(m),d0,DL(mr)); + ndl_copy(DL(tnm),DL(mr)); } } } @@ -4810,9 +5030,63 @@ ND ndv_mul_nmv_trunc(int mod,NMV m0,NDV p,UINT *d) 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; }