=================================================================== RCS file: /home/cvs/OpenXM_contrib2/asir2000/engine/nd.c,v retrieving revision 1.238 retrieving revision 1.240 diff -u -p -r1.238 -r1.240 --- OpenXM_contrib2/asir2000/engine/nd.c 2017/08/31 02:36:21 1.238 +++ OpenXM_contrib2/asir2000/engine/nd.c 2017/09/15 01:52:51 1.240 @@ -1,4 +1,4 @@ -/* $OpenXM: OpenXM_contrib2/asir2000/engine/nd.c,v 1.237 2017/04/09 02:23:12 noro Exp $ */ +/* $OpenXM: OpenXM_contrib2/asir2000/engine/nd.c,v 1.239 2017/09/14 01:34:53 noro Exp $ */ #include "nd.h" @@ -98,6 +98,10 @@ void Pox_cmo_rpc(NODE,Obj *); ND nd_add_lf(ND p1,ND p2); void nd_mul_c_lf(ND p,GZ mul); void ndv_mul_c_lf(NDV p,GZ mul); +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,ND_pairs *nz); +NODE nd_f4_red_mod64_main(int m,ND_pairs sp0,int nsp,UINT *s0vect,int col, + NM_ind_pair *rvect,int *rhead,IndArray *imat,int nred,ND_pairs *nz); NODE nd_f4_red_lf_main(int m,ND_pairs sp0,int nsp,int trace,UINT *s0vect,int col, NM_ind_pair *rvect,int *rhead,IndArray *imat,int nred); int nd_gauss_elim_lf(mpz_t **mat0,int *sugar,int row,int col,int *colstat); @@ -2909,6 +2913,18 @@ ND_pairs nd_minsugarp( ND_pairs d, ND_pairs *prest ) return dm0; } +int nd_tdeg(NDV c) +{ + int wmax = 0; + int i,len; + NMV a; + + len = LEN(c); + for ( a = BDY(c), i = 0; i < len; i++, NMV_ADV(a) ) + wmax = MAX(TD(DL(a)),wmax); + return wmax; +} + int ndv_newps(int m,NDV a,NDV aq,int f4) { int len; @@ -2941,11 +2957,21 @@ int ndv_newps(int m,NDV a,NDV aq,int f4) } else error("ndv_newps : invalud modulus"); nd_bound[nd_psn] = ndv_compute_bound(aq); - SG(r) = SG(aq); ndl_copy(HDL(aq),DL(r)); +#if 1 + SG(r) = SG(aq); +#else + SG(r) = nd_tdeg(aq); +#endif + ndl_copy(HDL(aq),DL(r)); } else { if ( !m ) register_hcf(a); nd_bound[nd_psn] = ndv_compute_bound(a); - SG(r) = SG(a); ndl_copy(HDL(a),DL(r)); +#if 1 + SG(r) = SG(a); +#else + SG(r) = nd_tdeg(a); +#endif + ndl_copy(HDL(a),DL(r)); if ( !m && !nd_vc ) nd_ps_gz[nd_psn] = ndvtondvgz(a); } if ( nd_demand ) { @@ -5796,6 +5822,27 @@ int nd_to_vect(int mod,UINT *s0,int n,ND d,UINT *r) return i; } +#if defined(__GNUC__) + +#define MOD128(a,c,m) ((a)=(((c)!=0||((a)>=(m)))?(((((U128)(c))<<64)+(a))%(m)):(a))) + +int nd_to_vect64(int mod,UINT *s0,int n,ND d,U64 *r) +{ + NM m; + UINT *t,*s; + int i; + + for ( i = 0; i < n; i++ ) r[i] = 0; + 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] = (U64)CM(m); + } + for ( i = 0; !r[i]; i++ ); + return i; +} +#endif + int nd_to_vect_q(UINT *s0,int n,ND d,Q *r) { NM m; @@ -6121,7 +6168,6 @@ int ndv_reduce_vect_gz(GZ *gvect,int trace,int col,Ind 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; @@ -6191,6 +6237,78 @@ int ndv_reduce_vect(int m,UINT *svect,int col,IndArray return maxrs; } +#if defined(__GNUC__) + +int ndv_reduce_vect64(int m,U64 *svect,U64 *cvect,int col,IndArray *imat,NM_ind_pair *rp0,int nred) +{ + int i,j,k,len,pos,prev; + U64 a,c,c1,c2; + IndArray ivect; + unsigned char *ivc; + unsigned short *ivs; + unsigned int *ivi; + NDV redv; + NMV mr; + NODE rp; + int maxrs; + + for ( i = 0; i < col; i++ ) cvect[i] = 0; + maxrs = 0; + for ( i = 0; i < nred; i++ ) { + ivect = imat[i]; + k = ivect->head; + a = svect[k]; c = cvect[k]; + MOD128(a,c,m); + svect[k] = a; cvect[k] = 0; + if ( c = svect[k] ) { + maxrs = MAX(maxrs,rp0[i]->sugar); + c = m-c; redv = nd_ps[rp0[i]->index]; + len = LEN(redv); mr = BDY(redv); + 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]; c1 = CM(mr); prev = pos; + if ( c1 ) { + c2 = svect[pos]+c1*c; + if ( c2 < svect[pos] ) cvect[pos]++; + svect[pos] = c2; + } + } + break; + case 2: + ivs = ivect->index.s; + for ( j = 1, NMV_ADV(mr); j < len; j++, NMV_ADV(mr) ) { + pos = prev+ivs[j]; c1 = CM(mr); prev = pos; + if ( c1 ) { + c2 = svect[pos]+c1*c; + if ( c2 < svect[pos] ) cvect[pos]++; + svect[pos] = c2; + } + } + break; + case 4: + ivi = ivect->index.i; + for ( j = 1, NMV_ADV(mr); j < len; j++, NMV_ADV(mr) ) { + pos = prev+ivi[j]; c1 = CM(mr); prev = pos; + if ( c1 ) { + c2 = svect[pos]+c1*c; + if ( c2 < svect[pos] ) cvect[pos]++; + svect[pos] = c2; + } + } + break; + } + } + } + for ( i = 0; i < col; i++ ) { + a = svect[i]; c = cvect[i]; MOD128(a,c,m); svect[i] = a; + } + return maxrs; +} +#endif + int ndv_reduce_vect_sf(int m,UINT *svect,int col,IndArray *imat,NM_ind_pair *rp0,int nred) { int i,j,k,len,pos,prev; @@ -6448,6 +6566,36 @@ NDV vect_to_ndv(UINT *vect,int spcol,int col,int *rhea } } +#if defined(__GNUC__) +NDV vect64_to_ndv(U64 *vect,int spcol,int col,int *rhead,UINT *s0vect) +{ + int j,k,len; + UINT *p; + UINT 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)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 ) + if ( !rhead[j] ) { + if ( c = (UINT)vect[k++] ) { + ndl_copy(p,DL(mr)); CM(mr) = c; NMV_ADV(mr); + } + } + MKNDV(nd_nvar,mr0,len,r); + return r; + } +} +#endif + NDV vect_to_ndv_2(unsigned long *vect,int col,UINT *s0vect) { int j,k,len; @@ -7079,8 +7227,14 @@ init_eg(&eg_search); imat[i] = nm_ind_pair_to_vect_compress(trace,s0vect,col,s0hash,rvect[i]); rhead[imat[i]->head] = 1; } - if ( m > 0 || m == -1 ) + if ( m > 0 ) +#if defined(__GNUC__) + r0 = nd_f4_red_mod64_main(m,sp0,nsp,s0vect,col,rvect,rhead,imat,nred,nz); +#else r0 = nd_f4_red_main(m,sp0,nsp,s0vect,col,rvect,rhead,imat,nred,nz); +#endif + else if ( m == -1 ) + r0 = nd_f4_red_sf_main(m,sp0,nsp,s0vect,col,rvect,rhead,imat,nred,nz); else if ( m == -2 ) r0 = nd_f4_red_lf_main(m,sp0,nsp,trace,s0vect,col,rvect,rhead,imat,nred); else @@ -7093,7 +7247,8 @@ init_eg(&eg_search); return r0; } -/* for small finite fields */ +/* for Fp, 2<=p<2^16 */ + 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,ND_pairs *nz) { @@ -7180,6 +7335,174 @@ NODE nd_f4_red_main(int m,ND_pairs sp0,int nsp,UINT *s return r0; } +#if defined(__GNUC__) +/* for Fp, 2^15=index.c); + + /* elimination (2nd step) */ + colstat = (int *)MALLOC(spcol*sizeof(int)); + rank = nd_gauss_elim_mod64(spmat,spsugar,spactive,sprow,spcol,m,colstat); + r0 = 0; + for ( i = 0; i < rank; i++ ) { + NEXTNODE(r0,r); BDY(r) = + (pointer)vect64_to_ndv(spmat[i],spcol,col,rhead,s0vect); + SG((NDV)BDY(r)) = spsugar[i]; + GCFREE(spmat[i]); + } + if ( r0 ) NEXT(r) = 0; + + for ( ; i < sprow; i++ ) GCFREE(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); + } + if ( nz ) { + for ( i = 0; i < rank-1; i++ ) NEXT(spactive[i]) = spactive[i+1]; + if ( rank > 0 ) { + NEXT(spactive[rank-1]) = 0; + *nz = spactive[0]; + } else + *nz = 0; + } + return r0; +} +#endif + +/* for small finite fields */ + +NODE nd_f4_red_sf_main(int m,ND_pairs sp0,int nsp,UINT *s0vect,int col, + NM_ind_pair *rvect,int *rhead,IndArray *imat,int nred,ND_pairs *nz) +{ + 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; + ND_pairs *spactive; + + spcol = col-nred; + get_eg(&eg0); + /* elimination (1st step) */ + spmat = (int **)MALLOC(nsp*sizeof(UINT *)); + svect = (UINT *)MALLOC(col*sizeof(UINT)); + spsugar = (int *)MALLOC(nsp*sizeof(int)); + spactive = !nz?0:(ND_pairs *)MALLOC(nsp*sizeof(ND_pairs)); + for ( a = sprow = 0, sp = sp0; a < nsp; a++, sp = NEXT(sp) ) { + nd_sp(m,0,sp,&spol); + if ( !spol ) continue; + nd_to_vect(m,s0vect,col,spol,svect); + maxrs = ndv_reduce_vect_sf(m,svect,col,imat,rvect,nred); + for ( i = 0; i < col; i++ ) if ( svect[i] ) break; + if ( i < col ) { + spmat[sprow] = v = (UINT *)MALLOC_ATOMIC(spcol*sizeof(UINT)); + for ( j = k = 0; j < col; j++ ) + if ( !rhead[j] ) v[k++] = svect[j]; + spsugar[sprow] = MAX(maxrs,SG(spol)); + if ( nz ) + spactive[sprow] = sp; + 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++ ) GCFREE(imat[i]->index.c); + + /* elimination (2nd step) */ + colstat = (int *)MALLOC(spcol*sizeof(int)); + rank = nd_gauss_elim_sf(spmat,spsugar,sprow,spcol,m,colstat); + r0 = 0; + for ( i = 0; i < rank; i++ ) { + NEXTNODE(r0,r); BDY(r) = + (pointer)vect_to_ndv(spmat[i],spcol,col,rhead,s0vect); + SG((NDV)BDY(r)) = spsugar[i]; + GCFREE(spmat[i]); + } + if ( r0 ) NEXT(r) = 0; + + for ( ; i < sprow; i++ ) GCFREE(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); + } + if ( nz ) { + for ( i = 0; i < rank-1; i++ ) NEXT(spactive[i]) = spactive[i+1]; + if ( rank > 0 ) { + NEXT(spactive[rank-1]) = 0; + *nz = spactive[0]; + } else + *nz = 0; + } + return r0; +} + NODE nd_f4_red_lf_main(int m,ND_pairs sp0,int nsp,int trace,UINT *s0vect,int col, NM_ind_pair *rvect,int *rhead,IndArray *imat,int nred) { @@ -7673,6 +7996,85 @@ int nd_gauss_elim_mod(int **mat0,int *sugar,ND_pairs * } return rank; } + +#if defined(__GNUC__) + +int nd_gauss_elim_mod64(U64 **mat,int *sugar,ND_pairs *spactive,int row,int col,int md,int *colstat) +{ + int i,j,k,l,rank,s; + U64 inv; + U64 a; + UINT c; + U64 *t,*pivot,*pk; + UINT *ck; + UINT **cmat; + UINT *ct; + ND_pairs pair; + + cmat = (UINT **)MALLOC(row*sizeof(UINT *)); + for ( i = 0; i < row; i++ ) { + cmat[i] = MALLOC_ATOMIC(col*sizeof(UINT)); + bzero(cmat[i],col*sizeof(UINT)); + } + + for ( rank = 0, j = 0; j < col; j++ ) { + for ( i = rank; i < row; i++ ) { + a = mat[i][j]; c = cmat[i][j]; + MOD128(a,c,md); + mat[i][j] = a; cmat[i][j] = 0; + } + for ( i = rank; i < row; i++ ) + if ( mat[i][j] ) + break; + if ( i == row ) { + colstat[j] = 0; + continue; + } else + colstat[j] = 1; + if ( i != rank ) { + t = mat[i]; mat[i] = mat[rank]; mat[rank] = t; + ct = cmat[i]; cmat[i] = cmat[rank]; cmat[rank] = ct; + s = sugar[i]; sugar[i] = sugar[rank]; sugar[rank] = s; + if ( spactive ) { + pair = spactive[i]; spactive[i] = spactive[rank]; + spactive[rank] = pair; + } + } + /* column j is normalized */ + s = sugar[rank]; + inv = invm((UINT)mat[rank][j],md); + /* normalize pivot row */ + for ( k = j, pk = mat[rank]+j, ck = cmat[rank]+j; k < col; k++, pk++, ck++ ) { + a = *pk; c = *ck; MOD128(a,c,md); *pk = (a*inv)%md; *ck = 0; + } + for ( i = rank+1; i < row; i++ ) { + if ( a = mat[i][j] ) { + sugar[i] = MAX(sugar[i],s); + red_by_vect64(md,mat[i]+j,cmat[i]+j,mat[rank]+j,(int)(md-a),col-j); + } + } + rank++; + } + for ( j = col-1, l = rank-1; j >= 0; j-- ) + if ( colstat[j] ) { + for ( k = j, pk = mat[l]+j, ck = cmat[l]+j; k < col; k++, pk++, ck++ ) { + a = *pk; c = *ck; MOD128(a,c,md); *pk = a; *ck = 0; + } + s = sugar[l]; + for ( i = 0; i < l; i++ ) { + a = mat[i][j]; c = cmat[i][j]; MOD128(a,c,md); mat[i][j] = a; cmat[i][j] = 0; + if ( a ) { + sugar[i] = MAX(sugar[i],s); + red_by_vect64(md,mat[i]+j,cmat[i]+j,mat[l]+j,(int)(md-a),col-j); + } + } + l--; + } + for ( i = 0; i < row; i++ ) GCFREE(cmat[i]); + GCFREE(cmat); + return rank; +} +#endif int nd_gauss_elim_sf(int **mat0,int *sugar,int row,int col,int md,int *colstat) {