1827bd09bSSatish Balay 2827bd09bSSatish Balay /*************************************xyt.c************************************ 3827bd09bSSatish Balay Module Name: xyt 4827bd09bSSatish Balay Module Info: 5827bd09bSSatish Balay 6827bd09bSSatish Balay author: Henry M. Tufo III 7827bd09bSSatish Balay e-mail: hmt@asci.uchicago.edu 8827bd09bSSatish Balay contact: 9827bd09bSSatish Balay +--------------------------------+--------------------------------+ 10827bd09bSSatish Balay |MCS Division - Building 221 |Department of Computer Science | 11827bd09bSSatish Balay |Argonne National Laboratory |Ryerson 152 | 12827bd09bSSatish Balay |9700 S. Cass Avenue |The University of Chicago | 13827bd09bSSatish Balay |Argonne, IL 60439 |Chicago, IL 60637 | 14827bd09bSSatish Balay |(630) 252-5354/5986 ph/fx |(773) 702-6019/8487 ph/fx | 15827bd09bSSatish Balay +--------------------------------+--------------------------------+ 16827bd09bSSatish Balay 17827bd09bSSatish Balay Last Modification: 3.20.01 18827bd09bSSatish Balay **************************************xyt.c***********************************/ 19c6db04a5SJed Brown #include <../src/ksp/pc/impls/tfs/tfs.h> 20827bd09bSSatish Balay 21827bd09bSSatish Balay #define LEFT -1 22827bd09bSSatish Balay #define RIGHT 1 23827bd09bSSatish Balay #define BOTH 0 24827bd09bSSatish Balay 25827bd09bSSatish Balay typedef struct xyt_solver_info { 2652f87cdaSBarry Smith PetscInt n, m, n_global, m_global; 2752f87cdaSBarry Smith PetscInt nnz, max_nnz, msg_buf_sz; 2852f87cdaSBarry Smith PetscInt *nsep, *lnsep, *fo, nfo, *stages; 2952f87cdaSBarry Smith PetscInt *xcol_sz, *xcol_indices; 30a501084fSBarry Smith PetscScalar **xcol_vals, *x, *solve_uu, *solve_w; 3152f87cdaSBarry Smith PetscInt *ycol_sz, *ycol_indices; 32a501084fSBarry Smith PetscScalar **ycol_vals, *y; 3352f87cdaSBarry Smith PetscInt nsolves; 34a501084fSBarry Smith PetscScalar tot_solve_time; 35827bd09bSSatish Balay } xyt_info; 36827bd09bSSatish Balay 37827bd09bSSatish Balay 38827bd09bSSatish Balay typedef struct matvec_info { 3952f87cdaSBarry Smith PetscInt n, m, n_global, m_global; 4052f87cdaSBarry Smith PetscInt *local2global; 41ca8e9878SJed Brown PCTFS_gs_ADT PCTFS_gs_handle; 42a501084fSBarry Smith PetscErrorCode (*matvec)(struct matvec_info*,PetscScalar*,PetscScalar*); 43827bd09bSSatish Balay void *grid_data; 44827bd09bSSatish Balay } mv_info; 45827bd09bSSatish Balay 46827bd09bSSatish Balay struct xyt_CDT{ 4752f87cdaSBarry Smith PetscInt id; 4852f87cdaSBarry Smith PetscInt ns; 4952f87cdaSBarry Smith PetscInt level; 50827bd09bSSatish Balay xyt_info *info; 51827bd09bSSatish Balay mv_info *mvi; 52827bd09bSSatish Balay }; 53827bd09bSSatish Balay 5452f87cdaSBarry Smith static PetscInt n_xyt=0; 5552f87cdaSBarry Smith static PetscInt n_xyt_handles=0; 56827bd09bSSatish Balay 57827bd09bSSatish Balay /* prototypes */ 583fdc5746SBarry Smith static PetscErrorCode do_xyt_solve(xyt_ADT xyt_handle, PetscScalar *rhs); 593fdc5746SBarry Smith static PetscErrorCode check_handle(xyt_ADT xyt_handle); 603fdc5746SBarry Smith static PetscErrorCode det_separators(xyt_ADT xyt_handle); 613fdc5746SBarry Smith static PetscErrorCode do_matvec(mv_info *A, PetscScalar *v, PetscScalar *u); 6252f87cdaSBarry Smith static PetscInt xyt_generate(xyt_ADT xyt_handle); 6352f87cdaSBarry Smith static PetscInt do_xyt_factor(xyt_ADT xyt_handle); 6452f87cdaSBarry Smith static mv_info *set_mvi(PetscInt *local2global, PetscInt n, PetscInt m, void *matvec, void *grid_data); 65827bd09bSSatish Balay 667b1ae94cSBarry Smith /**************************************xyt.c***********************************/ 677b1ae94cSBarry Smith xyt_ADT XYT_new(void) 68827bd09bSSatish Balay { 69827bd09bSSatish Balay xyt_ADT xyt_handle; 70827bd09bSSatish Balay 71827bd09bSSatish Balay /* rolling count on n_xyt ... pot. problem here */ 72827bd09bSSatish Balay n_xyt_handles++; 73a501084fSBarry Smith xyt_handle = (xyt_ADT)malloc(sizeof(struct xyt_CDT)); 74827bd09bSSatish Balay xyt_handle->id = ++n_xyt; 75827bd09bSSatish Balay xyt_handle->info = NULL; 76827bd09bSSatish Balay xyt_handle->mvi = NULL; 77827bd09bSSatish Balay 78827bd09bSSatish Balay return(xyt_handle); 79827bd09bSSatish Balay } 80827bd09bSSatish Balay 817b1ae94cSBarry Smith /**************************************xyt.c***********************************/ 8252f87cdaSBarry Smith PetscInt XYT_factor(xyt_ADT xyt_handle, /* prev. allocated xyt handle */ 8352f87cdaSBarry Smith PetscInt *local2global, /* global column mapping */ 8452f87cdaSBarry Smith PetscInt n, /* local num rows */ 8552f87cdaSBarry Smith PetscInt m, /* local num cols */ 86827bd09bSSatish Balay void *matvec, /* b_loc=A_local.x_loc */ 87827bd09bSSatish Balay void *grid_data /* grid data for matvec */ 88827bd09bSSatish Balay ) 89827bd09bSSatish Balay { 90827bd09bSSatish Balay 91b1c944f5SJed Brown PCTFS_comm_init(); 92827bd09bSSatish Balay check_handle(xyt_handle); 93827bd09bSSatish Balay 94827bd09bSSatish Balay /* only 2^k for now and all nodes participating */ 95b1c944f5SJed Brown if ((1<<(xyt_handle->level=PCTFS_i_log2_num_nodes))!=PCTFS_num_nodes) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"only 2^k for now and MPI_COMM_WORLD!!! %D != %D\n",1<<PCTFS_i_log2_num_nodes,PCTFS_num_nodes); 96827bd09bSSatish Balay 97827bd09bSSatish Balay /* space for X info */ 98a501084fSBarry Smith xyt_handle->info = (xyt_info*)malloc(sizeof(xyt_info)); 99827bd09bSSatish Balay 100827bd09bSSatish Balay /* set up matvec handles */ 101827bd09bSSatish Balay xyt_handle->mvi = set_mvi(local2global, n, m, matvec, grid_data); 102827bd09bSSatish Balay 103827bd09bSSatish Balay /* matrix is assumed to be of full rank */ 104827bd09bSSatish Balay /* LATER we can reset to indicate rank def. */ 105827bd09bSSatish Balay xyt_handle->ns=0; 106827bd09bSSatish Balay 107827bd09bSSatish Balay /* determine separators and generate firing order - NB xyt info set here */ 108827bd09bSSatish Balay det_separators(xyt_handle); 109827bd09bSSatish Balay 110827bd09bSSatish Balay return(do_xyt_factor(xyt_handle)); 111827bd09bSSatish Balay } 112827bd09bSSatish Balay 1137b1ae94cSBarry Smith /**************************************xyt.c***********************************/ 1148cda6cd7SBarry Smith PetscInt XYT_solve(xyt_ADT xyt_handle, PetscScalar *x, PetscScalar *b) 115827bd09bSSatish Balay { 116b1c944f5SJed Brown PCTFS_comm_init(); 117827bd09bSSatish Balay check_handle(xyt_handle); 118827bd09bSSatish Balay 119827bd09bSSatish Balay /* need to copy b into x? */ 120827bd09bSSatish Balay if (b) 121ca8e9878SJed Brown {PCTFS_rvec_copy(x,b,xyt_handle->mvi->n);} 122827bd09bSSatish Balay do_xyt_solve(xyt_handle,x); 123827bd09bSSatish Balay 124827bd09bSSatish Balay return(0); 125827bd09bSSatish Balay } 126827bd09bSSatish Balay 1277b1ae94cSBarry Smith /**************************************xyt.c***********************************/ 12852f87cdaSBarry Smith PetscInt XYT_free(xyt_ADT xyt_handle) 129827bd09bSSatish Balay { 130b1c944f5SJed Brown PCTFS_comm_init(); 131827bd09bSSatish Balay check_handle(xyt_handle); 132827bd09bSSatish Balay n_xyt_handles--; 133827bd09bSSatish Balay 134a501084fSBarry Smith free(xyt_handle->info->nsep); 135a501084fSBarry Smith free(xyt_handle->info->lnsep); 136a501084fSBarry Smith free(xyt_handle->info->fo); 137a501084fSBarry Smith free(xyt_handle->info->stages); 138a501084fSBarry Smith free(xyt_handle->info->solve_uu); 139a501084fSBarry Smith free(xyt_handle->info->solve_w); 140a501084fSBarry Smith free(xyt_handle->info->x); 141a501084fSBarry Smith free(xyt_handle->info->xcol_vals); 142a501084fSBarry Smith free(xyt_handle->info->xcol_sz); 143a501084fSBarry Smith free(xyt_handle->info->xcol_indices); 144a501084fSBarry Smith free(xyt_handle->info->y); 145a501084fSBarry Smith free(xyt_handle->info->ycol_vals); 146a501084fSBarry Smith free(xyt_handle->info->ycol_sz); 147a501084fSBarry Smith free(xyt_handle->info->ycol_indices); 148a501084fSBarry Smith free(xyt_handle->info); 149a501084fSBarry Smith free(xyt_handle->mvi->local2global); 150ca8e9878SJed Brown PCTFS_gs_free(xyt_handle->mvi->PCTFS_gs_handle); 151a501084fSBarry Smith free(xyt_handle->mvi); 152a501084fSBarry Smith free(xyt_handle); 153827bd09bSSatish Balay 154827bd09bSSatish Balay 155827bd09bSSatish Balay /* if the check fails we nuke */ 156a501084fSBarry Smith /* if NULL pointer passed to free we nuke */ 157827bd09bSSatish Balay /* if the calls to free fail that's not my problem */ 158827bd09bSSatish Balay return(0); 159827bd09bSSatish Balay } 160827bd09bSSatish Balay 1617b1ae94cSBarry Smith /**************************************xyt.c***********************************/ 16252f87cdaSBarry Smith PetscInt XYT_stats(xyt_ADT xyt_handle) 163827bd09bSSatish Balay { 16452f87cdaSBarry Smith PetscInt op[] = {NON_UNIFORM,GL_MIN,GL_MAX,GL_ADD,GL_MIN,GL_MAX,GL_ADD,GL_MIN,GL_MAX,GL_ADD}; 16552f87cdaSBarry Smith PetscInt fop[] = {NON_UNIFORM,GL_MIN,GL_MAX,GL_ADD}; 16652f87cdaSBarry Smith PetscInt vals[9], work[9]; 167a501084fSBarry Smith PetscScalar fvals[3], fwork[3]; 168827bd09bSSatish Balay 169b1c944f5SJed Brown PCTFS_comm_init(); 170827bd09bSSatish Balay check_handle(xyt_handle); 171827bd09bSSatish Balay 172827bd09bSSatish Balay /* if factorization not done there are no stats */ 1737d0a6c19SBarry Smith if (!xyt_handle->info||!xyt_handle->mvi) { 174b1c944f5SJed Brown if (!PCTFS_my_id) PetscPrintf(PETSC_COMM_WORLD,"XYT_stats() :: no stats available!\n"); 175827bd09bSSatish Balay return 1; 176827bd09bSSatish Balay } 177827bd09bSSatish Balay 178827bd09bSSatish Balay vals[0]=vals[1]=vals[2]=xyt_handle->info->nnz; 179827bd09bSSatish Balay vals[3]=vals[4]=vals[5]=xyt_handle->mvi->n; 180827bd09bSSatish Balay vals[6]=vals[7]=vals[8]=xyt_handle->info->msg_buf_sz; 181b1c944f5SJed Brown PCTFS_giop(vals,work,sizeof(op)/sizeof(op[0])-1,op); 182827bd09bSSatish Balay 183827bd09bSSatish Balay fvals[0]=fvals[1]=fvals[2] 184827bd09bSSatish Balay =xyt_handle->info->tot_solve_time/xyt_handle->info->nsolves++; 185b1c944f5SJed Brown PCTFS_grop(fvals,fwork,sizeof(fop)/sizeof(fop[0])-1,fop); 186827bd09bSSatish Balay 187b1c944f5SJed Brown if (!PCTFS_my_id) { 188b1c944f5SJed Brown PetscPrintf(PETSC_COMM_WORLD,"%D :: min xyt_nnz=%D\n",PCTFS_my_id,vals[0]); 189b1c944f5SJed Brown PetscPrintf(PETSC_COMM_WORLD,"%D :: max xyt_nnz=%D\n",PCTFS_my_id,vals[1]); 190b1c944f5SJed Brown PetscPrintf(PETSC_COMM_WORLD,"%D :: avg xyt_nnz=%g\n",PCTFS_my_id,1.0*vals[2]/PCTFS_num_nodes); 191b1c944f5SJed Brown PetscPrintf(PETSC_COMM_WORLD,"%D :: tot xyt_nnz=%D\n",PCTFS_my_id,vals[2]); 192b1c944f5SJed Brown PetscPrintf(PETSC_COMM_WORLD,"%D :: xyt C(2d) =%g\n",PCTFS_my_id,vals[2]/(pow(1.0*vals[5],1.5))); 193b1c944f5SJed Brown PetscPrintf(PETSC_COMM_WORLD,"%D :: xyt C(3d) =%g\n",PCTFS_my_id,vals[2]/(pow(1.0*vals[5],1.6667))); 194b1c944f5SJed Brown PetscPrintf(PETSC_COMM_WORLD,"%D :: min xyt_n =%D\n",PCTFS_my_id,vals[3]); 195b1c944f5SJed Brown PetscPrintf(PETSC_COMM_WORLD,"%D :: max xyt_n =%D\n",PCTFS_my_id,vals[4]); 196b1c944f5SJed Brown PetscPrintf(PETSC_COMM_WORLD,"%D :: avg xyt_n =%g\n",PCTFS_my_id,1.0*vals[5]/PCTFS_num_nodes); 197b1c944f5SJed Brown PetscPrintf(PETSC_COMM_WORLD,"%D :: tot xyt_n =%D\n",PCTFS_my_id,vals[5]); 198b1c944f5SJed Brown PetscPrintf(PETSC_COMM_WORLD,"%D :: min xyt_buf=%D\n",PCTFS_my_id,vals[6]); 199b1c944f5SJed Brown PetscPrintf(PETSC_COMM_WORLD,"%D :: max xyt_buf=%D\n",PCTFS_my_id,vals[7]); 200b1c944f5SJed Brown PetscPrintf(PETSC_COMM_WORLD,"%D :: avg xyt_buf=%g\n",PCTFS_my_id,1.0*vals[8]/PCTFS_num_nodes); 201b1c944f5SJed Brown PetscPrintf(PETSC_COMM_WORLD,"%D :: min xyt_slv=%g\n",PCTFS_my_id,fvals[0]); 202b1c944f5SJed Brown PetscPrintf(PETSC_COMM_WORLD,"%D :: max xyt_slv=%g\n",PCTFS_my_id,fvals[1]); 203b1c944f5SJed Brown PetscPrintf(PETSC_COMM_WORLD,"%D :: avg xyt_slv=%g\n",PCTFS_my_id,fvals[2]/PCTFS_num_nodes); 204827bd09bSSatish Balay } 205827bd09bSSatish Balay 206827bd09bSSatish Balay return(0); 207827bd09bSSatish Balay } 208827bd09bSSatish Balay 209827bd09bSSatish Balay 210827bd09bSSatish Balay /*************************************xyt.c************************************ 211827bd09bSSatish Balay 212827bd09bSSatish Balay Description: get A_local, local portion of global coarse matrix which 213827bd09bSSatish Balay is a row dist. nxm matrix w/ n<m. 214827bd09bSSatish Balay o my_ml holds address of ML struct associated w/A_local and coarse grid 215827bd09bSSatish Balay o local2global holds global number of column i (i=0,...,m-1) 216827bd09bSSatish Balay o local2global holds global number of row i (i=0,...,n-1) 217827bd09bSSatish Balay o mylocmatvec performs A_local . vec_local (note that gs is performed using 218ca8e9878SJed Brown PCTFS_gs_init/gop). 219827bd09bSSatish Balay 220827bd09bSSatish Balay mylocmatvec = my_ml->Amat[grid_tag].matvec->external; 221827bd09bSSatish Balay mylocmatvec (void :: void *data, double *in, double *out) 222827bd09bSSatish Balay **************************************xyt.c***********************************/ 22352f87cdaSBarry Smith static PetscInt do_xyt_factor(xyt_ADT xyt_handle) 224827bd09bSSatish Balay { 2257b1ae94cSBarry Smith return xyt_generate(xyt_handle); 226827bd09bSSatish Balay } 227827bd09bSSatish Balay 2287b1ae94cSBarry Smith /**************************************xyt.c***********************************/ 22952f87cdaSBarry Smith static PetscInt xyt_generate(xyt_ADT xyt_handle) 230827bd09bSSatish Balay { 23152f87cdaSBarry Smith PetscInt i,j,k,idx; 23252f87cdaSBarry Smith PetscInt dim, col; 233a501084fSBarry Smith PetscScalar *u, *uu, *v, *z, *w, alpha, alpha_w; 23452f87cdaSBarry Smith PetscInt *segs; 23552f87cdaSBarry Smith PetscInt op[] = {GL_ADD,0}; 23652f87cdaSBarry Smith PetscInt off, len; 237a501084fSBarry Smith PetscScalar *x_ptr, *y_ptr; 23852f87cdaSBarry Smith PetscInt *iptr, flag; 23952f87cdaSBarry Smith PetscInt start=0, end, work; 24052f87cdaSBarry Smith PetscInt op2[] = {GL_MIN,0}; 241ca8e9878SJed Brown PCTFS_gs_ADT PCTFS_gs_handle; 24252f87cdaSBarry Smith PetscInt *nsep, *lnsep, *fo; 24352f87cdaSBarry Smith PetscInt a_n=xyt_handle->mvi->n; 24452f87cdaSBarry Smith PetscInt a_m=xyt_handle->mvi->m; 24552f87cdaSBarry Smith PetscInt *a_local2global=xyt_handle->mvi->local2global; 24652f87cdaSBarry Smith PetscInt level; 24752f87cdaSBarry Smith PetscInt n, m; 24852f87cdaSBarry Smith PetscInt *xcol_sz, *xcol_indices, *stages; 249a501084fSBarry Smith PetscScalar **xcol_vals, *x; 25052f87cdaSBarry Smith PetscInt *ycol_sz, *ycol_indices; 251a501084fSBarry Smith PetscScalar **ycol_vals, *y; 25252f87cdaSBarry Smith PetscInt n_global; 25352f87cdaSBarry Smith PetscInt xt_nnz=0, xt_max_nnz=0; 25452f87cdaSBarry Smith PetscInt yt_nnz=0, yt_max_nnz=0; 25552f87cdaSBarry Smith PetscInt xt_zero_nnz =0; 25652f87cdaSBarry Smith PetscInt xt_zero_nnz_0=0; 25752f87cdaSBarry Smith PetscInt yt_zero_nnz =0; 25852f87cdaSBarry Smith PetscInt yt_zero_nnz_0=0; 2594a0de3f6SBarry Smith PetscBLASInt i1 = 1,dlen; 260a501084fSBarry Smith PetscScalar dm1 = -1.0; 261d1528f56SBarry Smith PetscErrorCode ierr; 262827bd09bSSatish Balay 263827bd09bSSatish Balay n=xyt_handle->mvi->n; 264827bd09bSSatish Balay nsep=xyt_handle->info->nsep; 265827bd09bSSatish Balay lnsep=xyt_handle->info->lnsep; 266827bd09bSSatish Balay fo=xyt_handle->info->fo; 267827bd09bSSatish Balay end=lnsep[0]; 268827bd09bSSatish Balay level=xyt_handle->level; 269ca8e9878SJed Brown PCTFS_gs_handle=xyt_handle->mvi->PCTFS_gs_handle; 270827bd09bSSatish Balay 271827bd09bSSatish Balay /* is there a null space? */ 272827bd09bSSatish Balay /* LATER add in ability to detect null space by checking alpha */ 273*db4deed7SKarl Rupp for (i=0, j=0; i<=level; i++) { j+=nsep[i]; } 274827bd09bSSatish Balay 275827bd09bSSatish Balay m = j-xyt_handle->ns; 27622d28d08SBarry Smith if (m!=j) { 27722d28d08SBarry Smith ierr = PetscPrintf(PETSC_COMM_WORLD,"xyt_generate() :: null space exists %D %D %D\n",m,j,xyt_handle->ns);CHKERRQ(ierr); 27822d28d08SBarry Smith } 279827bd09bSSatish Balay 2804a0de3f6SBarry Smith ierr = PetscInfo2(0,"xyt_generate() :: X(%D,%D)\n",n,m);CHKERRQ(ierr); 281827bd09bSSatish Balay 282827bd09bSSatish Balay /* get and initialize storage for x local */ 283827bd09bSSatish Balay /* note that x local is nxm and stored by columns */ 28452f87cdaSBarry Smith xcol_sz = (PetscInt*) malloc(m*sizeof(PetscInt)); 28552f87cdaSBarry Smith xcol_indices = (PetscInt*) malloc((2*m+1)*sizeof(PetscInt)); 286a501084fSBarry Smith xcol_vals = (PetscScalar **) malloc(m*sizeof(PetscScalar *)); 287*db4deed7SKarl Rupp for (i=j=0; i<m; i++, j+=2) { 288827bd09bSSatish Balay xcol_indices[j]=xcol_indices[j+1]=xcol_sz[i]=-1; 289827bd09bSSatish Balay xcol_vals[i] = NULL; 290827bd09bSSatish Balay } 291827bd09bSSatish Balay xcol_indices[j]=-1; 292827bd09bSSatish Balay 293827bd09bSSatish Balay /* get and initialize storage for y local */ 294827bd09bSSatish Balay /* note that y local is nxm and stored by columns */ 29552f87cdaSBarry Smith ycol_sz = (PetscInt*) malloc(m*sizeof(PetscInt)); 29652f87cdaSBarry Smith ycol_indices = (PetscInt*) malloc((2*m+1)*sizeof(PetscInt)); 297a501084fSBarry Smith ycol_vals = (PetscScalar **) malloc(m*sizeof(PetscScalar *)); 298*db4deed7SKarl Rupp for (i=j=0; i<m; i++, j+=2) { 299827bd09bSSatish Balay ycol_indices[j]=ycol_indices[j+1]=ycol_sz[i]=-1; 300827bd09bSSatish Balay ycol_vals[i] = NULL; 301827bd09bSSatish Balay } 302827bd09bSSatish Balay ycol_indices[j]=-1; 303827bd09bSSatish Balay 304827bd09bSSatish Balay /* size of separators for each sub-hc working from bottom of tree to top */ 305827bd09bSSatish Balay /* this looks like nsep[]=segments */ 30652f87cdaSBarry Smith stages = (PetscInt*) malloc((level+1)*sizeof(PetscInt)); 30752f87cdaSBarry Smith segs = (PetscInt*) malloc((level+1)*sizeof(PetscInt)); 308ca8e9878SJed Brown PCTFS_ivec_zero(stages,level+1); 309ca8e9878SJed Brown PCTFS_ivec_copy(segs,nsep,level+1); 310*db4deed7SKarl Rupp for (i=0; i<level; i++) { segs[i+1] += segs[i]; } 311827bd09bSSatish Balay stages[0] = segs[0]; 312827bd09bSSatish Balay 313827bd09bSSatish Balay /* temporary vectors */ 314a501084fSBarry Smith u = (PetscScalar *) malloc(n*sizeof(PetscScalar)); 315a501084fSBarry Smith z = (PetscScalar *) malloc(n*sizeof(PetscScalar)); 316a501084fSBarry Smith v = (PetscScalar *) malloc(a_m*sizeof(PetscScalar)); 317a501084fSBarry Smith uu = (PetscScalar *) malloc(m*sizeof(PetscScalar)); 318a501084fSBarry Smith w = (PetscScalar *) malloc(m*sizeof(PetscScalar)); 319827bd09bSSatish Balay 320827bd09bSSatish Balay /* extra nnz due to replication of vertices across separators */ 321*db4deed7SKarl Rupp for (i=1, j=0; i<=level; i++) { j+=nsep[i]; } 322827bd09bSSatish Balay 323827bd09bSSatish Balay /* storage for sparse x values */ 324827bd09bSSatish Balay n_global = xyt_handle->info->n_global; 325b1c944f5SJed Brown xt_max_nnz = yt_max_nnz = (PetscInt)(2.5*pow(1.0*n_global,1.6667) + j*n/2)/PCTFS_num_nodes; 326a501084fSBarry Smith x = (PetscScalar *) malloc(xt_max_nnz*sizeof(PetscScalar)); 327a501084fSBarry Smith y = (PetscScalar *) malloc(yt_max_nnz*sizeof(PetscScalar)); 328827bd09bSSatish Balay 329827bd09bSSatish Balay /* LATER - can embed next sep to fire in gs */ 330827bd09bSSatish Balay /* time to make the donuts - generate X factor */ 331*db4deed7SKarl Rupp for (dim=i=j=0;i<m;i++) { 332827bd09bSSatish Balay /* time to move to the next level? */ 333d4af0d30SBarry Smith while (i==segs[dim]) { 334e32f2f54SBarry Smith if (dim==level) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"dim about to exceed level\n"); 335827bd09bSSatish Balay stages[dim++]=i; 336827bd09bSSatish Balay end+=lnsep[dim]; 337827bd09bSSatish Balay } 338827bd09bSSatish Balay stages[dim]=i; 339827bd09bSSatish Balay 340827bd09bSSatish Balay /* which column are we firing? */ 341827bd09bSSatish Balay /* i.e. set v_l */ 342827bd09bSSatish Balay /* use new seps and do global min across hc to determine which one to fire */ 343827bd09bSSatish Balay (start<end) ? (col=fo[start]) : (col=INT_MAX); 344b1c944f5SJed Brown PCTFS_giop_hc(&col,&work,1,op2,dim); 345827bd09bSSatish Balay 346827bd09bSSatish Balay /* shouldn't need this */ 347*db4deed7SKarl Rupp if (col==INT_MAX) { 348f1ed62a8SBarry Smith ierr = PetscInfo(0,"hey ... col==INT_MAX??\n");CHKERRQ(ierr); 349827bd09bSSatish Balay continue; 350827bd09bSSatish Balay } 351827bd09bSSatish Balay 352827bd09bSSatish Balay /* do I own it? I should */ 353ca8e9878SJed Brown PCTFS_rvec_zero(v ,a_m); 354827bd09bSSatish Balay if (col==fo[start]) 355827bd09bSSatish Balay { 356827bd09bSSatish Balay start++; 357ca8e9878SJed Brown idx=PCTFS_ivec_linear_search(col, a_local2global, a_n); 358*db4deed7SKarl Rupp if (idx!=-1) { v[idx] = 1.0; j++; } 359c1235816SBarry Smith else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"NOT FOUND!\n"); 360*db4deed7SKarl Rupp } else { 361ca8e9878SJed Brown idx=PCTFS_ivec_linear_search(col, a_local2global, a_m); 362*db4deed7SKarl Rupp if (idx!=-1) {v[idx] = 1.0;} 363827bd09bSSatish Balay } 364827bd09bSSatish Balay 365827bd09bSSatish Balay /* perform u = A.v_l */ 366ca8e9878SJed Brown PCTFS_rvec_zero(u,n); 367827bd09bSSatish Balay do_matvec(xyt_handle->mvi,v,u); 368827bd09bSSatish Balay 369827bd09bSSatish Balay /* uu = X^T.u_l (local portion) */ 370827bd09bSSatish Balay /* technically only need to zero out first i entries */ 371827bd09bSSatish Balay /* later turn this into an XYT_solve call ? */ 372ca8e9878SJed Brown PCTFS_rvec_zero(uu,m); 373827bd09bSSatish Balay y_ptr=y; 374827bd09bSSatish Balay iptr = ycol_indices; 375*db4deed7SKarl Rupp for (k=0; k<i; k++) { 376827bd09bSSatish Balay off = *iptr++; 377827bd09bSSatish Balay len = *iptr++; 3780805154bSBarry Smith dlen = PetscBLASIntCast(len); 3794a0de3f6SBarry Smith uu[k] = BLASdot_(&dlen,u+off,&i1,y_ptr,&i1); 380827bd09bSSatish Balay y_ptr+=len; 381827bd09bSSatish Balay } 382827bd09bSSatish Balay 383827bd09bSSatish Balay /* uu = X^T.u_l (comm portion) */ 384b1c944f5SJed Brown PCTFS_ssgl_radd (uu, w, dim, stages); 385827bd09bSSatish Balay 386827bd09bSSatish Balay /* z = X.uu */ 387ca8e9878SJed Brown PCTFS_rvec_zero(z,n); 388827bd09bSSatish Balay x_ptr=x; 389827bd09bSSatish Balay iptr = xcol_indices; 390*db4deed7SKarl Rupp for (k=0; k<i; k++) { 391827bd09bSSatish Balay off = *iptr++; 392827bd09bSSatish Balay len = *iptr++; 3930805154bSBarry Smith dlen = PetscBLASIntCast(len); 3944a0de3f6SBarry Smith BLASaxpy_(&dlen,&uu[k],x_ptr,&i1,z+off,&i1); 395827bd09bSSatish Balay x_ptr+=len; 396827bd09bSSatish Balay } 397827bd09bSSatish Balay 398827bd09bSSatish Balay /* compute v_l = v_l - z */ 399ca8e9878SJed Brown PCTFS_rvec_zero(v+a_n,a_m-a_n); 4000805154bSBarry Smith dlen = PetscBLASIntCast(n); 4014a0de3f6SBarry Smith BLASaxpy_(&dlen,&dm1,z,&i1,v,&i1); 402827bd09bSSatish Balay 403827bd09bSSatish Balay /* compute u_l = A.v_l */ 404*db4deed7SKarl Rupp if (a_n!=a_m) { PCTFS_gs_gop_hc(PCTFS_gs_handle,v,"+\0",dim); } 405ca8e9878SJed Brown PCTFS_rvec_zero(u,n); 406827bd09bSSatish Balay do_matvec(xyt_handle->mvi,v,u); 407827bd09bSSatish Balay 408827bd09bSSatish Balay /* compute sqrt(alpha) = sqrt(u_l^T.u_l) - local portion */ 4090805154bSBarry Smith dlen = PetscBLASIntCast(n); 4104a0de3f6SBarry Smith alpha = BLASdot_(&dlen,u,&i1,u,&i1); 411827bd09bSSatish Balay /* compute sqrt(alpha) = sqrt(u_l^T.u_l) - comm portion */ 412b1c944f5SJed Brown PCTFS_grop_hc(&alpha, &alpha_w, 1, op, dim); 413827bd09bSSatish Balay 4148f1a2a5eSBarry Smith alpha = (PetscScalar) PetscSqrtReal((PetscReal)alpha); 415827bd09bSSatish Balay 416827bd09bSSatish Balay /* check for small alpha */ 417827bd09bSSatish Balay /* LATER use this to detect and determine null space */ 418e32f2f54SBarry Smith if (fabs(alpha)<1.0e-14) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_PLIB,"bad alpha! %g\n",alpha); 419827bd09bSSatish Balay 420827bd09bSSatish Balay /* compute v_l = v_l/sqrt(alpha) */ 421ca8e9878SJed Brown PCTFS_rvec_scale(v,1.0/alpha,n); 422ca8e9878SJed Brown PCTFS_rvec_scale(u,1.0/alpha,n); 423827bd09bSSatish Balay 424827bd09bSSatish Balay /* add newly generated column, v_l, to X */ 425827bd09bSSatish Balay flag = 1; 426827bd09bSSatish Balay off=len=0; 427*db4deed7SKarl Rupp for (k=0; k<n; k++) { 428*db4deed7SKarl Rupp if (v[k]!=0.0) { 429827bd09bSSatish Balay len=k; 430*db4deed7SKarl Rupp if (flag) {off=k; flag=0;} 431827bd09bSSatish Balay } 432827bd09bSSatish Balay } 433827bd09bSSatish Balay 434827bd09bSSatish Balay len -= (off-1); 435827bd09bSSatish Balay 436*db4deed7SKarl Rupp if (len>0) { 437*db4deed7SKarl Rupp if ((xt_nnz+len)>xt_max_nnz) { 438f1ed62a8SBarry Smith ierr = PetscInfo(0,"increasing space for X by 2x!\n");CHKERRQ(ierr); 439827bd09bSSatish Balay xt_max_nnz *= 2; 440a501084fSBarry Smith x_ptr = (PetscScalar *) malloc(xt_max_nnz*sizeof(PetscScalar)); 441ca8e9878SJed Brown PCTFS_rvec_copy(x_ptr,x,xt_nnz); 442a501084fSBarry Smith free(x); 443827bd09bSSatish Balay x = x_ptr; 444827bd09bSSatish Balay x_ptr+=xt_nnz; 445827bd09bSSatish Balay } 446827bd09bSSatish Balay xt_nnz += len; 447ca8e9878SJed Brown PCTFS_rvec_copy(x_ptr,v+off,len); 448827bd09bSSatish Balay 449827bd09bSSatish Balay /* keep track of number of zeros */ 450*db4deed7SKarl Rupp if (dim) { 451*db4deed7SKarl Rupp for (k=0; k<len; k++) { 452*db4deed7SKarl Rupp if (x_ptr[k]==0.0) {xt_zero_nnz++;} 453827bd09bSSatish Balay } 454*db4deed7SKarl Rupp } else { 455*db4deed7SKarl Rupp for (k=0; k<len; k++) { 456*db4deed7SKarl Rupp if (x_ptr[k]==0.0) { xt_zero_nnz_0++; } 457827bd09bSSatish Balay } 458827bd09bSSatish Balay } 459827bd09bSSatish Balay xcol_indices[2*i] = off; 460827bd09bSSatish Balay xcol_sz[i] = xcol_indices[2*i+1] = len; 461827bd09bSSatish Balay xcol_vals[i] = x_ptr; 462*db4deed7SKarl Rupp } else { 463827bd09bSSatish Balay xcol_indices[2*i] = 0; 464827bd09bSSatish Balay xcol_sz[i] = xcol_indices[2*i+1] = 0; 465827bd09bSSatish Balay xcol_vals[i] = x_ptr; 466827bd09bSSatish Balay } 467827bd09bSSatish Balay 468827bd09bSSatish Balay 469827bd09bSSatish Balay /* add newly generated column, u_l, to Y */ 470827bd09bSSatish Balay flag = 1; 471827bd09bSSatish Balay off=len=0; 472*db4deed7SKarl Rupp for (k=0; k<n; k++) { 473*db4deed7SKarl Rupp if (u[k]!=0.0) { 474827bd09bSSatish Balay len=k; 475*db4deed7SKarl Rupp if (flag) { off=k; flag=0; } 476827bd09bSSatish Balay } 477827bd09bSSatish Balay } 478827bd09bSSatish Balay 479827bd09bSSatish Balay len -= (off-1); 480827bd09bSSatish Balay 481*db4deed7SKarl Rupp if (len>0) { 482*db4deed7SKarl Rupp if ((yt_nnz+len)>yt_max_nnz) { 483f1ed62a8SBarry Smith ierr = PetscInfo(0,"increasing space for Y by 2x!\n");CHKERRQ(ierr); 484827bd09bSSatish Balay yt_max_nnz *= 2; 485a501084fSBarry Smith y_ptr = (PetscScalar *) malloc(yt_max_nnz*sizeof(PetscScalar)); 486ca8e9878SJed Brown PCTFS_rvec_copy(y_ptr,y,yt_nnz); 487a501084fSBarry Smith free(y); 488827bd09bSSatish Balay y = y_ptr; 489827bd09bSSatish Balay y_ptr+=yt_nnz; 490827bd09bSSatish Balay } 491827bd09bSSatish Balay yt_nnz += len; 492ca8e9878SJed Brown PCTFS_rvec_copy(y_ptr,u+off,len); 493827bd09bSSatish Balay 494827bd09bSSatish Balay /* keep track of number of zeros */ 495*db4deed7SKarl Rupp if (dim) { 496*db4deed7SKarl Rupp for (k=0; k<len; k++) { 497*db4deed7SKarl Rupp if (y_ptr[k]==0.0) { yt_zero_nnz++; } 498827bd09bSSatish Balay } 499*db4deed7SKarl Rupp } else { 500*db4deed7SKarl Rupp for (k=0; k<len; k++) { 501*db4deed7SKarl Rupp if (y_ptr[k]==0.0) { yt_zero_nnz_0++; } 502827bd09bSSatish Balay } 503827bd09bSSatish Balay } 504827bd09bSSatish Balay ycol_indices[2*i] = off; 505827bd09bSSatish Balay ycol_sz[i] = ycol_indices[2*i+1] = len; 506827bd09bSSatish Balay ycol_vals[i] = y_ptr; 507*db4deed7SKarl Rupp } else { 508827bd09bSSatish Balay ycol_indices[2*i] = 0; 509827bd09bSSatish Balay ycol_sz[i] = ycol_indices[2*i+1] = 0; 510827bd09bSSatish Balay ycol_vals[i] = y_ptr; 511827bd09bSSatish Balay } 512827bd09bSSatish Balay } 513827bd09bSSatish Balay 514827bd09bSSatish Balay /* close off stages for execution phase */ 515*db4deed7SKarl Rupp while (dim!=level) { 516827bd09bSSatish Balay stages[dim++]=i; 5174a0de3f6SBarry Smith ierr = PetscInfo2(0,"disconnected!!! dim(%D)!=level(%D)\n",dim,level);CHKERRQ(ierr); 518827bd09bSSatish Balay } 519827bd09bSSatish Balay stages[dim]=i; 520827bd09bSSatish Balay 521827bd09bSSatish Balay xyt_handle->info->n=xyt_handle->mvi->n; 522827bd09bSSatish Balay xyt_handle->info->m=m; 523827bd09bSSatish Balay xyt_handle->info->nnz=xt_nnz + yt_nnz; 524827bd09bSSatish Balay xyt_handle->info->max_nnz=xt_max_nnz + yt_max_nnz; 525827bd09bSSatish Balay xyt_handle->info->msg_buf_sz=stages[level]-stages[0]; 526a501084fSBarry Smith xyt_handle->info->solve_uu = (PetscScalar *) malloc(m*sizeof(PetscScalar)); 527a501084fSBarry Smith xyt_handle->info->solve_w = (PetscScalar *) malloc(m*sizeof(PetscScalar)); 528827bd09bSSatish Balay xyt_handle->info->x=x; 529827bd09bSSatish Balay xyt_handle->info->xcol_vals=xcol_vals; 530827bd09bSSatish Balay xyt_handle->info->xcol_sz=xcol_sz; 531827bd09bSSatish Balay xyt_handle->info->xcol_indices=xcol_indices; 532827bd09bSSatish Balay xyt_handle->info->stages=stages; 533827bd09bSSatish Balay xyt_handle->info->y=y; 534827bd09bSSatish Balay xyt_handle->info->ycol_vals=ycol_vals; 535827bd09bSSatish Balay xyt_handle->info->ycol_sz=ycol_sz; 536827bd09bSSatish Balay xyt_handle->info->ycol_indices=ycol_indices; 537827bd09bSSatish Balay 538a501084fSBarry Smith free(segs); 539a501084fSBarry Smith free(u); 540a501084fSBarry Smith free(v); 541a501084fSBarry Smith free(uu); 542a501084fSBarry Smith free(z); 543a501084fSBarry Smith free(w); 544827bd09bSSatish Balay 545827bd09bSSatish Balay return(0); 546827bd09bSSatish Balay } 547827bd09bSSatish Balay 5487b1ae94cSBarry Smith /**************************************xyt.c***********************************/ 5490924e98cSBarry Smith static PetscErrorCode do_xyt_solve(xyt_ADT xyt_handle, PetscScalar *uc) 550827bd09bSSatish Balay { 55152f87cdaSBarry Smith PetscInt off, len, *iptr; 55252f87cdaSBarry Smith PetscInt level =xyt_handle->level; 55352f87cdaSBarry Smith PetscInt n =xyt_handle->info->n; 55452f87cdaSBarry Smith PetscInt m =xyt_handle->info->m; 55552f87cdaSBarry Smith PetscInt *stages =xyt_handle->info->stages; 55652f87cdaSBarry Smith PetscInt *xcol_indices=xyt_handle->info->xcol_indices; 55752f87cdaSBarry Smith PetscInt *ycol_indices=xyt_handle->info->ycol_indices; 558a501084fSBarry Smith PetscScalar *x_ptr, *y_ptr, *uu_ptr; 559a501084fSBarry Smith PetscScalar *solve_uu=xyt_handle->info->solve_uu; 560a501084fSBarry Smith PetscScalar *solve_w =xyt_handle->info->solve_w; 561a501084fSBarry Smith PetscScalar *x =xyt_handle->info->x; 562a501084fSBarry Smith PetscScalar *y =xyt_handle->info->y; 5634a0de3f6SBarry Smith PetscBLASInt i1 = 1,dlen; 564827bd09bSSatish Balay 5650924e98cSBarry Smith PetscFunctionBegin; 566827bd09bSSatish Balay uu_ptr=solve_uu; 567ca8e9878SJed Brown PCTFS_rvec_zero(uu_ptr,m); 568827bd09bSSatish Balay 569827bd09bSSatish Balay /* x = X.Y^T.b */ 570827bd09bSSatish Balay /* uu = Y^T.b */ 571827bd09bSSatish Balay for (y_ptr=y,iptr=ycol_indices; *iptr!=-1; y_ptr+=len) 572827bd09bSSatish Balay { 573827bd09bSSatish Balay off=*iptr++; len=*iptr++; 5740805154bSBarry Smith dlen = PetscBLASIntCast(len); 5754a0de3f6SBarry Smith *uu_ptr++ = BLASdot_(&dlen,uc+off,&i1,y_ptr,&i1); 576827bd09bSSatish Balay } 577827bd09bSSatish Balay 578827bd09bSSatish Balay /* comunication of beta */ 579827bd09bSSatish Balay uu_ptr=solve_uu; 580b1c944f5SJed Brown if (level) { PCTFS_ssgl_radd(uu_ptr, solve_w, level, stages); } 581827bd09bSSatish Balay 582ca8e9878SJed Brown PCTFS_rvec_zero(uc,n); 583827bd09bSSatish Balay 584827bd09bSSatish Balay /* x = X.uu */ 585*db4deed7SKarl Rupp for (x_ptr=x,iptr=xcol_indices; *iptr!=-1; x_ptr+=len) { 586827bd09bSSatish Balay off=*iptr++; len=*iptr++; 5870805154bSBarry Smith dlen = PetscBLASIntCast(len); 5884a0de3f6SBarry Smith BLASaxpy_(&dlen,uu_ptr++,x_ptr,&i1,uc+off,&i1); 589827bd09bSSatish Balay } 5900924e98cSBarry Smith PetscFunctionReturn(0); 591827bd09bSSatish Balay } 592827bd09bSSatish Balay 5937b1ae94cSBarry Smith /**************************************xyt.c***********************************/ 5940924e98cSBarry Smith static PetscErrorCode check_handle(xyt_ADT xyt_handle) 595827bd09bSSatish Balay { 59652f87cdaSBarry Smith PetscInt vals[2], work[2], op[] = {NON_UNIFORM,GL_MIN,GL_MAX}; 597827bd09bSSatish Balay 5980924e98cSBarry Smith PetscFunctionBegin; 599e32f2f54SBarry Smith if (!xyt_handle) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_PLIB,"check_handle() :: bad handle :: NULL %D\n",xyt_handle); 600827bd09bSSatish Balay 601827bd09bSSatish Balay vals[0]=vals[1]=xyt_handle->id; 602b1c944f5SJed Brown PCTFS_giop(vals,work,sizeof(op)/sizeof(op[0])-1,op); 603e32f2f54SBarry Smith if ((vals[0]!=vals[1])||(xyt_handle->id<=0)) SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_PLIB,"check_handle() :: bad handle :: id mismatch min/max %D/%D %D\n", vals[0],vals[1], xyt_handle->id); 6040924e98cSBarry Smith PetscFunctionReturn(0); 605827bd09bSSatish Balay } 606827bd09bSSatish Balay 6077b1ae94cSBarry Smith /**************************************xyt.c***********************************/ 6080924e98cSBarry Smith static PetscErrorCode det_separators(xyt_ADT xyt_handle) 609827bd09bSSatish Balay { 61052f87cdaSBarry Smith PetscInt i, ct, id; 61152f87cdaSBarry Smith PetscInt mask, edge, *iptr; 61252f87cdaSBarry Smith PetscInt *dir, *used; 61352f87cdaSBarry Smith PetscInt sum[4], w[4]; 614a501084fSBarry Smith PetscScalar rsum[4], rw[4]; 61552f87cdaSBarry Smith PetscInt op[] = {GL_ADD,0}; 616a501084fSBarry Smith PetscScalar *lhs, *rhs; 61752f87cdaSBarry Smith PetscInt *nsep, *lnsep, *fo, nfo=0; 618ca8e9878SJed Brown PCTFS_gs_ADT PCTFS_gs_handle=xyt_handle->mvi->PCTFS_gs_handle; 61952f87cdaSBarry Smith PetscInt *local2global=xyt_handle->mvi->local2global; 62052f87cdaSBarry Smith PetscInt n=xyt_handle->mvi->n; 62152f87cdaSBarry Smith PetscInt m=xyt_handle->mvi->m; 62252f87cdaSBarry Smith PetscInt level=xyt_handle->level; 623ab824b78SBarry Smith PetscInt shared=0; 624d1528f56SBarry Smith PetscErrorCode ierr; 625827bd09bSSatish Balay 6260924e98cSBarry Smith PetscFunctionBegin; 62752f87cdaSBarry Smith dir = (PetscInt*)malloc(sizeof(PetscInt)*(level+1)); 62852f87cdaSBarry Smith nsep = (PetscInt*)malloc(sizeof(PetscInt)*(level+1)); 62952f87cdaSBarry Smith lnsep= (PetscInt*)malloc(sizeof(PetscInt)*(level+1)); 63052f87cdaSBarry Smith fo = (PetscInt*)malloc(sizeof(PetscInt)*(n+1)); 63152f87cdaSBarry Smith used = (PetscInt*)malloc(sizeof(PetscInt)*n); 632827bd09bSSatish Balay 633ca8e9878SJed Brown PCTFS_ivec_zero(dir ,level+1); 634ca8e9878SJed Brown PCTFS_ivec_zero(nsep ,level+1); 635ca8e9878SJed Brown PCTFS_ivec_zero(lnsep,level+1); 636ca8e9878SJed Brown PCTFS_ivec_set (fo ,-1,n+1); 637ca8e9878SJed Brown PCTFS_ivec_zero(used,n); 638827bd09bSSatish Balay 6398cda6cd7SBarry Smith lhs = (PetscScalar*)malloc(sizeof(PetscScalar)*m); 6408cda6cd7SBarry Smith rhs = (PetscScalar*)malloc(sizeof(PetscScalar)*m); 641827bd09bSSatish Balay 642827bd09bSSatish Balay /* determine the # of unique dof */ 643ca8e9878SJed Brown PCTFS_rvec_zero(lhs,m); 644ca8e9878SJed Brown PCTFS_rvec_set(lhs,1.0,n); 645ca8e9878SJed Brown PCTFS_gs_gop_hc(PCTFS_gs_handle,lhs,"+\0",level); 646ca8e9878SJed Brown ierr = PetscInfo(0,"done first PCTFS_gs_gop_hc\n");CHKERRQ(ierr); 647ca8e9878SJed Brown PCTFS_rvec_zero(rsum,2); 648827bd09bSSatish Balay for (ct=i=0;i<n;i++) 649827bd09bSSatish Balay { 650*db4deed7SKarl Rupp if (lhs[i]!=0.0) { rsum[0]+=1.0/lhs[i]; rsum[1]+=lhs[i]; } 651827bd09bSSatish Balay 652*db4deed7SKarl Rupp if (lhs[i]!=1.0) { 653ab824b78SBarry Smith shared=1; 654827bd09bSSatish Balay } 655827bd09bSSatish Balay } 656827bd09bSSatish Balay 657b1c944f5SJed Brown PCTFS_grop_hc(rsum,rw,2,op,level); 658827bd09bSSatish Balay rsum[0]+=0.1; 659827bd09bSSatish Balay rsum[1]+=0.1; 660827bd09bSSatish Balay 66152f87cdaSBarry Smith xyt_handle->info->n_global=xyt_handle->info->m_global=(PetscInt) rsum[0]; 66252f87cdaSBarry Smith xyt_handle->mvi->n_global =xyt_handle->mvi->m_global =(PetscInt) rsum[0]; 663827bd09bSSatish Balay 664827bd09bSSatish Balay /* determine separator sets top down */ 665827bd09bSSatish Balay if (shared) 666827bd09bSSatish Balay { 667827bd09bSSatish Balay /* solution is to do as in the symmetric shared case but then */ 668827bd09bSSatish Balay /* pick the sub-hc with the most free dofs and do a mat-vec */ 669827bd09bSSatish Balay /* and pick up the responses on the other sub-hc from the */ 670827bd09bSSatish Balay /* initial separator set obtained from the symm. shared case */ 671e32f2f54SBarry Smith SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"shared dof separator determination not ready ... see hmt!!!\n"); 672*db4deed7SKarl Rupp for (iptr=fo+n,id=PCTFS_my_id,mask=PCTFS_num_nodes>>1,edge=level;edge>0;edge--,mask>>=1) { 673*db4deed7SKarl Rupp 674827bd09bSSatish Balay /* set rsh of hc, fire, and collect lhs responses */ 675ca8e9878SJed Brown (id<mask) ? PCTFS_rvec_zero(lhs,m) : PCTFS_rvec_set(lhs,1.0,m); 676ca8e9878SJed Brown PCTFS_gs_gop_hc(PCTFS_gs_handle,lhs,"+\0",edge); 677827bd09bSSatish Balay 678827bd09bSSatish Balay /* set lsh of hc, fire, and collect rhs responses */ 679ca8e9878SJed Brown (id<mask) ? PCTFS_rvec_set(rhs,1.0,m) : PCTFS_rvec_zero(rhs,m); 680ca8e9878SJed Brown PCTFS_gs_gop_hc(PCTFS_gs_handle,rhs,"+\0",edge); 681827bd09bSSatish Balay 682*db4deed7SKarl Rupp for (i=0;i<n;i++) { 683*db4deed7SKarl Rupp if (id< mask) { 684*db4deed7SKarl Rupp if (lhs[i]!=0.0) { lhs[i]=1.0; } 685827bd09bSSatish Balay } 686*db4deed7SKarl Rupp if (id>=mask) { 687*db4deed7SKarl Rupp if (rhs[i]!=0.0) { rhs[i]=1.0; } 688827bd09bSSatish Balay } 689827bd09bSSatish Balay } 690827bd09bSSatish Balay 691*db4deed7SKarl Rupp if (id< mask) { PCTFS_gs_gop_hc(PCTFS_gs_handle,lhs,"+\0",edge-1); } 692*db4deed7SKarl Rupp else { PCTFS_gs_gop_hc(PCTFS_gs_handle,rhs,"+\0",edge-1); } 693827bd09bSSatish Balay 694827bd09bSSatish Balay /* count number of dofs I own that have signal and not in sep set */ 695ca8e9878SJed Brown PCTFS_rvec_zero(rsum,4); 696*db4deed7SKarl Rupp for (PCTFS_ivec_zero(sum,4),ct=i=0;i<n;i++) { 697*db4deed7SKarl Rupp if (!used[i]) { 698827bd09bSSatish Balay /* number of unmarked dofs on node */ 699827bd09bSSatish Balay ct++; 700827bd09bSSatish Balay /* number of dofs to be marked on lhs hc */ 701*db4deed7SKarl Rupp if (id< mask) { 702*db4deed7SKarl Rupp if (lhs[i]!=0.0) { sum[0]++; rsum[0]+=1.0/lhs[i]; } 703827bd09bSSatish Balay } 704827bd09bSSatish Balay /* number of dofs to be marked on rhs hc */ 705*db4deed7SKarl Rupp if (id>=mask) { 706*db4deed7SKarl Rupp if (rhs[i]!=0.0) { sum[1]++; rsum[1]+=1.0/rhs[i]; } 707827bd09bSSatish Balay } 708827bd09bSSatish Balay } 709827bd09bSSatish Balay } 710827bd09bSSatish Balay 711827bd09bSSatish Balay /* go for load balance - choose half with most unmarked dofs, bias LHS */ 712827bd09bSSatish Balay (id<mask) ? (sum[2]=ct) : (sum[3]=ct); 713827bd09bSSatish Balay (id<mask) ? (rsum[2]=ct) : (rsum[3]=ct); 714b1c944f5SJed Brown PCTFS_giop_hc(sum,w,4,op,edge); 715b1c944f5SJed Brown PCTFS_grop_hc(rsum,rw,4,op,edge); 716827bd09bSSatish Balay rsum[0]+=0.1; rsum[1]+=0.1; rsum[2]+=0.1; rsum[3]+=0.1; 717827bd09bSSatish Balay 718*db4deed7SKarl Rupp if (id<mask) { 719827bd09bSSatish Balay /* mark dofs I own that have signal and not in sep set */ 720*db4deed7SKarl Rupp for (ct=i=0;i<n;i++) { 721*db4deed7SKarl Rupp if ((!used[i])&&(lhs[i]!=0.0)) { 722827bd09bSSatish Balay ct++; nfo++; 723827bd09bSSatish Balay 724c1235816SBarry Smith if (nfo>n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"nfo about to exceed n\n"); 725827bd09bSSatish Balay 726827bd09bSSatish Balay *--iptr = local2global[i]; 727827bd09bSSatish Balay used[i]=edge; 728827bd09bSSatish Balay } 729827bd09bSSatish Balay } 730ca8e9878SJed Brown if (ct>1) {PCTFS_ivec_sort(iptr,ct);} 731827bd09bSSatish Balay 732827bd09bSSatish Balay lnsep[edge]=ct; 73352f87cdaSBarry Smith nsep[edge]=(PetscInt) rsum[0]; 734827bd09bSSatish Balay dir [edge]=LEFT; 735827bd09bSSatish Balay } 736827bd09bSSatish Balay 737*db4deed7SKarl Rupp if (id>=mask) { 738827bd09bSSatish Balay /* mark dofs I own that have signal and not in sep set */ 739*db4deed7SKarl Rupp for (ct=i=0;i<n;i++) { 740*db4deed7SKarl Rupp if ((!used[i])&&(rhs[i]!=0.0)) { 741827bd09bSSatish Balay ct++; nfo++; 742827bd09bSSatish Balay 743c1235816SBarry Smith if (nfo>n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"nfo about to exceed n\n"); 744827bd09bSSatish Balay 745827bd09bSSatish Balay *--iptr = local2global[i]; 746827bd09bSSatish Balay used[i]=edge; 747827bd09bSSatish Balay } 748827bd09bSSatish Balay } 749ca8e9878SJed Brown if (ct>1) { PCTFS_ivec_sort(iptr,ct); } 750827bd09bSSatish Balay 751827bd09bSSatish Balay lnsep[edge]=ct; 75252f87cdaSBarry Smith nsep[edge]= (PetscInt) rsum[1]; 753827bd09bSSatish Balay dir [edge]=RIGHT; 754827bd09bSSatish Balay } 755827bd09bSSatish Balay 756827bd09bSSatish Balay /* LATER or we can recur on these to order seps at this level */ 757827bd09bSSatish Balay /* do we need full set of separators for this? */ 758827bd09bSSatish Balay 759827bd09bSSatish Balay /* fold rhs hc into lower */ 760*db4deed7SKarl Rupp if (id>=mask) { id-=mask; } 761827bd09bSSatish Balay } 762*db4deed7SKarl Rupp } else { 763*db4deed7SKarl Rupp for (iptr=fo+n,id=PCTFS_my_id,mask=PCTFS_num_nodes>>1,edge=level;edge>0;edge--,mask>>=1) { 764827bd09bSSatish Balay /* set rsh of hc, fire, and collect lhs responses */ 765ca8e9878SJed Brown (id<mask) ? PCTFS_rvec_zero(lhs,m) : PCTFS_rvec_set(lhs,1.0,m); 766ca8e9878SJed Brown PCTFS_gs_gop_hc(PCTFS_gs_handle,lhs,"+\0",edge); 767827bd09bSSatish Balay 768827bd09bSSatish Balay /* set lsh of hc, fire, and collect rhs responses */ 769ca8e9878SJed Brown (id<mask) ? PCTFS_rvec_set(rhs,1.0,m) : PCTFS_rvec_zero(rhs,m); 770ca8e9878SJed Brown PCTFS_gs_gop_hc(PCTFS_gs_handle,rhs,"+\0",edge); 771827bd09bSSatish Balay 772827bd09bSSatish Balay /* count number of dofs I own that have signal and not in sep set */ 773*db4deed7SKarl Rupp for (PCTFS_ivec_zero(sum,4),ct=i=0;i<n;i++) { 774*db4deed7SKarl Rupp if (!used[i]) { 775827bd09bSSatish Balay /* number of unmarked dofs on node */ 776827bd09bSSatish Balay ct++; 777827bd09bSSatish Balay /* number of dofs to be marked on lhs hc */ 778827bd09bSSatish Balay if ((id< mask)&&(lhs[i]!=0.0)) { sum[0]++; } 779827bd09bSSatish Balay /* number of dofs to be marked on rhs hc */ 780827bd09bSSatish Balay if ((id>=mask)&&(rhs[i]!=0.0)) { sum[1]++; } 781827bd09bSSatish Balay } 782827bd09bSSatish Balay } 783827bd09bSSatish Balay 784827bd09bSSatish Balay /* for the non-symmetric case we need separators of width 2 */ 785827bd09bSSatish Balay /* so take both sides */ 786827bd09bSSatish Balay (id<mask) ? (sum[2]=ct) : (sum[3]=ct); 787b1c944f5SJed Brown PCTFS_giop_hc(sum,w,4,op,edge); 788827bd09bSSatish Balay 789827bd09bSSatish Balay ct=0; 790*db4deed7SKarl Rupp if (id<mask) { 791827bd09bSSatish Balay /* mark dofs I own that have signal and not in sep set */ 792*db4deed7SKarl Rupp for (i=0;i<n;i++) { 793*db4deed7SKarl Rupp if ((!used[i])&&(lhs[i]!=0.0)) { 794827bd09bSSatish Balay ct++; nfo++; 795827bd09bSSatish Balay *--iptr = local2global[i]; 796827bd09bSSatish Balay used[i]=edge; 797827bd09bSSatish Balay } 798827bd09bSSatish Balay } 799827bd09bSSatish Balay /* LSH hc summation of ct should be sum[0] */ 800*db4deed7SKarl Rupp } else { 801827bd09bSSatish Balay /* mark dofs I own that have signal and not in sep set */ 802*db4deed7SKarl Rupp for (i=0;i<n;i++) { 803*db4deed7SKarl Rupp if ((!used[i])&&(rhs[i]!=0.0)) { 804827bd09bSSatish Balay ct++; nfo++; 805827bd09bSSatish Balay *--iptr = local2global[i]; 806827bd09bSSatish Balay used[i]=edge; 807827bd09bSSatish Balay } 808827bd09bSSatish Balay } 809827bd09bSSatish Balay /* RSH hc summation of ct should be sum[1] */ 810827bd09bSSatish Balay } 811827bd09bSSatish Balay 812ca8e9878SJed Brown if (ct>1) { PCTFS_ivec_sort(iptr,ct); } 813827bd09bSSatish Balay lnsep[edge]=ct; 814827bd09bSSatish Balay nsep[edge]=sum[0]+sum[1]; 815827bd09bSSatish Balay dir [edge]=BOTH; 816827bd09bSSatish Balay 817827bd09bSSatish Balay /* LATER or we can recur on these to order seps at this level */ 818827bd09bSSatish Balay /* do we need full set of separators for this? */ 819827bd09bSSatish Balay 820827bd09bSSatish Balay /* fold rhs hc into lower */ 821827bd09bSSatish Balay if (id>=mask) 822827bd09bSSatish Balay {id-=mask;} 823827bd09bSSatish Balay } 824827bd09bSSatish Balay } 825827bd09bSSatish Balay 826827bd09bSSatish Balay /* level 0 is on processor case - so mark the remainder */ 827*db4deed7SKarl Rupp for (ct=i=0;i<n;i++) { 828*db4deed7SKarl Rupp if (!used[i]) { 829827bd09bSSatish Balay ct++; nfo++; 830827bd09bSSatish Balay *--iptr = local2global[i]; 831827bd09bSSatish Balay used[i]=edge; 832827bd09bSSatish Balay } 833827bd09bSSatish Balay } 834ca8e9878SJed Brown if (ct>1) {PCTFS_ivec_sort(iptr,ct);} 835827bd09bSSatish Balay lnsep[edge]=ct; 836827bd09bSSatish Balay nsep [edge]=ct; 837827bd09bSSatish Balay dir [edge]=BOTH; 838827bd09bSSatish Balay 839827bd09bSSatish Balay xyt_handle->info->nsep=nsep; 840827bd09bSSatish Balay xyt_handle->info->lnsep=lnsep; 841827bd09bSSatish Balay xyt_handle->info->fo=fo; 842827bd09bSSatish Balay xyt_handle->info->nfo=nfo; 843827bd09bSSatish Balay 844a501084fSBarry Smith free(dir); 845a501084fSBarry Smith free(lhs); 846a501084fSBarry Smith free(rhs); 847a501084fSBarry Smith free(used); 8480924e98cSBarry Smith PetscFunctionReturn(0); 849827bd09bSSatish Balay } 850827bd09bSSatish Balay 8517b1ae94cSBarry Smith /**************************************xyt.c***********************************/ 85252f87cdaSBarry Smith static mv_info *set_mvi(PetscInt *local2global, PetscInt n, PetscInt m, void *matvec, void *grid_data) 853827bd09bSSatish Balay { 854827bd09bSSatish Balay mv_info *mvi; 855827bd09bSSatish Balay 856827bd09bSSatish Balay 857a501084fSBarry Smith mvi = (mv_info*)malloc(sizeof(mv_info)); 858827bd09bSSatish Balay mvi->n=n; 859827bd09bSSatish Balay mvi->m=m; 860827bd09bSSatish Balay mvi->n_global=-1; 861827bd09bSSatish Balay mvi->m_global=-1; 86252f87cdaSBarry Smith mvi->local2global=(PetscInt*)malloc((m+1)*sizeof(PetscInt)); 863ca8e9878SJed Brown PCTFS_ivec_copy(mvi->local2global,local2global,m); 864827bd09bSSatish Balay mvi->local2global[m] = INT_MAX; 865a501084fSBarry Smith mvi->matvec=(PetscErrorCode (*)(mv_info*,PetscScalar*,PetscScalar*))matvec; 866827bd09bSSatish Balay mvi->grid_data=grid_data; 867827bd09bSSatish Balay 868827bd09bSSatish Balay /* set xyt communication handle to perform restricted matvec */ 869ca8e9878SJed Brown mvi->PCTFS_gs_handle = PCTFS_gs_init(local2global, m, PCTFS_num_nodes); 870827bd09bSSatish Balay 871827bd09bSSatish Balay return(mvi); 872827bd09bSSatish Balay } 873827bd09bSSatish Balay 8747b1ae94cSBarry Smith /**************************************xyt.c***********************************/ 8753fdc5746SBarry Smith static PetscErrorCode do_matvec(mv_info *A, PetscScalar *v, PetscScalar *u) 876827bd09bSSatish Balay { 8770924e98cSBarry Smith PetscFunctionBegin; 878827bd09bSSatish Balay A->matvec((mv_info*)A->grid_data,v,u); 8790924e98cSBarry Smith PetscFunctionReturn(0); 880827bd09bSSatish Balay } 881827bd09bSSatish Balay 882827bd09bSSatish Balay 883827bd09bSSatish Balay 884