xref: /petsc/src/ksp/pc/impls/tfs/xyt.c (revision 2c71b3e237ead271e4f3aa1505f92bf476e3413d)
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 typedef struct matvec_info {
3852f87cdaSBarry Smith   PetscInt     n, m, n_global, m_global;
3952f87cdaSBarry Smith   PetscInt     *local2global;
40ca8e9878SJed Brown   PCTFS_gs_ADT PCTFS_gs_handle;
41a501084fSBarry Smith   PetscErrorCode (*matvec)(struct matvec_info*,PetscScalar*,PetscScalar*);
42827bd09bSSatish Balay   void *grid_data;
43827bd09bSSatish Balay } mv_info;
44827bd09bSSatish Balay 
45827bd09bSSatish Balay struct xyt_CDT {
4652f87cdaSBarry Smith   PetscInt id;
4752f87cdaSBarry Smith   PetscInt ns;
4852f87cdaSBarry Smith   PetscInt level;
49827bd09bSSatish Balay   xyt_info *info;
50827bd09bSSatish Balay   mv_info  *mvi;
51827bd09bSSatish Balay };
52827bd09bSSatish Balay 
5352f87cdaSBarry Smith static PetscInt n_xyt        =0;
5452f87cdaSBarry Smith static PetscInt n_xyt_handles=0;
55827bd09bSSatish Balay 
56827bd09bSSatish Balay /* prototypes */
573fdc5746SBarry Smith static PetscErrorCode do_xyt_solve(xyt_ADT xyt_handle, PetscScalar *rhs);
583fdc5746SBarry Smith static PetscErrorCode check_handle(xyt_ADT xyt_handle);
593fdc5746SBarry Smith static PetscErrorCode det_separators(xyt_ADT xyt_handle);
603fdc5746SBarry Smith static PetscErrorCode do_matvec(mv_info *A, PetscScalar *v, PetscScalar *u);
6138dcbb09SBarry Smith static PetscErrorCode xyt_generate(xyt_ADT xyt_handle);
6238dcbb09SBarry Smith static PetscErrorCode do_xyt_factor(xyt_ADT xyt_handle);
635c8f6a95SKarl Rupp static mv_info *set_mvi(PetscInt *local2global, PetscInt n, PetscInt m, PetscErrorCode (*matvec)(mv_info*,PetscScalar*,PetscScalar*), void *grid_data);
64827bd09bSSatish Balay 
657b1ae94cSBarry Smith /**************************************xyt.c***********************************/
667b1ae94cSBarry Smith xyt_ADT XYT_new(void)
67827bd09bSSatish Balay {
68827bd09bSSatish Balay   xyt_ADT xyt_handle;
69827bd09bSSatish Balay 
70827bd09bSSatish Balay   /* rolling count on n_xyt ... pot. problem here */
71827bd09bSSatish Balay   n_xyt_handles++;
72a501084fSBarry Smith   xyt_handle       = (xyt_ADT)malloc(sizeof(struct xyt_CDT));
73827bd09bSSatish Balay   xyt_handle->id   = ++n_xyt;
74827bd09bSSatish Balay   xyt_handle->info = NULL;
75827bd09bSSatish Balay   xyt_handle->mvi  = NULL;
76827bd09bSSatish Balay 
77827bd09bSSatish Balay   return(xyt_handle);
78827bd09bSSatish Balay }
79827bd09bSSatish Balay 
807b1ae94cSBarry Smith /**************************************xyt.c***********************************/
8138dcbb09SBarry Smith PetscErrorCode XYT_factor(xyt_ADT xyt_handle,     /* prev. allocated xyt  handle */
8252f87cdaSBarry Smith                     PetscInt *local2global, /* global column mapping       */
8352f87cdaSBarry Smith                     PetscInt n,             /* local num rows              */
8452f87cdaSBarry Smith                     PetscInt m,             /* local num cols              */
855c8f6a95SKarl Rupp                     PetscErrorCode (*matvec)(void*,PetscScalar*,PetscScalar*), /* b_loc=A_local.x_loc         */
861147fc2aSKarl Rupp                     void *grid_data)        /* grid data for matvec        */
87827bd09bSSatish Balay {
88827bd09bSSatish Balay 
89b1c944f5SJed Brown   PCTFS_comm_init();
90827bd09bSSatish Balay   check_handle(xyt_handle);
91827bd09bSSatish Balay 
92827bd09bSSatish Balay   /* only 2^k for now and all nodes participating */
93*2c71b3e2SJacob Faibussowitsch   PetscCheckFalse((1<<(xyt_handle->level=PCTFS_i_log2_num_nodes))!=PCTFS_num_nodes,PETSC_COMM_SELF,PETSC_ERR_PLIB,"only 2^k for now and MPI_COMM_WORLD!!! %D != %D",1<<PCTFS_i_log2_num_nodes,PCTFS_num_nodes);
94827bd09bSSatish Balay 
95827bd09bSSatish Balay   /* space for X info */
96a501084fSBarry Smith   xyt_handle->info = (xyt_info*)malloc(sizeof(xyt_info));
97827bd09bSSatish Balay 
98827bd09bSSatish Balay   /* set up matvec handles */
995c8f6a95SKarl Rupp   xyt_handle->mvi = set_mvi(local2global, n, m, (PetscErrorCode (*)(mv_info*,PetscScalar*,PetscScalar*))matvec, grid_data);
100827bd09bSSatish Balay 
101827bd09bSSatish Balay   /* matrix is assumed to be of full rank */
102827bd09bSSatish Balay   /* LATER we can reset to indicate rank def. */
103827bd09bSSatish Balay   xyt_handle->ns=0;
104827bd09bSSatish Balay 
105827bd09bSSatish Balay   /* determine separators and generate firing order - NB xyt info set here */
106827bd09bSSatish Balay   det_separators(xyt_handle);
107827bd09bSSatish Balay 
108827bd09bSSatish Balay   return(do_xyt_factor(xyt_handle));
109827bd09bSSatish Balay }
110827bd09bSSatish Balay 
1117b1ae94cSBarry Smith /**************************************xyt.c***********************************/
11238dcbb09SBarry Smith PetscErrorCode XYT_solve(xyt_ADT xyt_handle, PetscScalar *x, PetscScalar *b)
113827bd09bSSatish Balay {
114b1c944f5SJed Brown   PCTFS_comm_init();
115827bd09bSSatish Balay   check_handle(xyt_handle);
116827bd09bSSatish Balay 
117827bd09bSSatish Balay   /* need to copy b into x? */
11859bc5b24SSatish Balay   if (b) PCTFS_rvec_copy(x,b,xyt_handle->mvi->n);
11938dcbb09SBarry Smith   return do_xyt_solve(xyt_handle,x);
120827bd09bSSatish Balay }
121827bd09bSSatish Balay 
1227b1ae94cSBarry Smith /**************************************xyt.c***********************************/
12338dcbb09SBarry Smith PetscErrorCode XYT_free(xyt_ADT xyt_handle)
124827bd09bSSatish Balay {
125b1c944f5SJed Brown   PCTFS_comm_init();
126827bd09bSSatish Balay   check_handle(xyt_handle);
127827bd09bSSatish Balay   n_xyt_handles--;
128827bd09bSSatish Balay 
129a501084fSBarry Smith   free(xyt_handle->info->nsep);
130a501084fSBarry Smith   free(xyt_handle->info->lnsep);
131a501084fSBarry Smith   free(xyt_handle->info->fo);
132a501084fSBarry Smith   free(xyt_handle->info->stages);
133a501084fSBarry Smith   free(xyt_handle->info->solve_uu);
134a501084fSBarry Smith   free(xyt_handle->info->solve_w);
135a501084fSBarry Smith   free(xyt_handle->info->x);
136a501084fSBarry Smith   free(xyt_handle->info->xcol_vals);
137a501084fSBarry Smith   free(xyt_handle->info->xcol_sz);
138a501084fSBarry Smith   free(xyt_handle->info->xcol_indices);
139a501084fSBarry Smith   free(xyt_handle->info->y);
140a501084fSBarry Smith   free(xyt_handle->info->ycol_vals);
141a501084fSBarry Smith   free(xyt_handle->info->ycol_sz);
142a501084fSBarry Smith   free(xyt_handle->info->ycol_indices);
143a501084fSBarry Smith   free(xyt_handle->info);
144a501084fSBarry Smith   free(xyt_handle->mvi->local2global);
145ca8e9878SJed Brown   PCTFS_gs_free(xyt_handle->mvi->PCTFS_gs_handle);
146a501084fSBarry Smith   free(xyt_handle->mvi);
147a501084fSBarry Smith   free(xyt_handle);
148827bd09bSSatish Balay 
149827bd09bSSatish Balay   /* if the check fails we nuke */
150a501084fSBarry Smith   /* if NULL pointer passed to free we nuke */
151827bd09bSSatish Balay   /* if the calls to free fail that's not my problem */
152827bd09bSSatish Balay   return(0);
153827bd09bSSatish Balay }
154827bd09bSSatish Balay 
1557b1ae94cSBarry Smith /**************************************xyt.c***********************************/
15638dcbb09SBarry Smith PetscErrorCode XYT_stats(xyt_ADT xyt_handle)
157827bd09bSSatish Balay {
15852f87cdaSBarry Smith   PetscInt    op[]  = {NON_UNIFORM,GL_MIN,GL_MAX,GL_ADD,GL_MIN,GL_MAX,GL_ADD,GL_MIN,GL_MAX,GL_ADD};
15952f87cdaSBarry Smith   PetscInt    fop[] = {NON_UNIFORM,GL_MIN,GL_MAX,GL_ADD};
16052f87cdaSBarry Smith   PetscInt    vals[9],  work[9];
161a501084fSBarry Smith   PetscScalar fvals[3], fwork[3];
162827bd09bSSatish Balay 
163b1c944f5SJed Brown   PCTFS_comm_init();
164827bd09bSSatish Balay   check_handle(xyt_handle);
165827bd09bSSatish Balay 
166827bd09bSSatish Balay   /* if factorization not done there are no stats */
1677d0a6c19SBarry Smith   if (!xyt_handle->info||!xyt_handle->mvi) {
168b1c944f5SJed Brown     if (!PCTFS_my_id) PetscPrintf(PETSC_COMM_WORLD,"XYT_stats() :: no stats available!\n");
169827bd09bSSatish Balay     return 1;
170827bd09bSSatish Balay   }
171827bd09bSSatish Balay 
172827bd09bSSatish Balay   vals[0]=vals[1]=vals[2]=xyt_handle->info->nnz;
173827bd09bSSatish Balay   vals[3]=vals[4]=vals[5]=xyt_handle->mvi->n;
174827bd09bSSatish Balay   vals[6]=vals[7]=vals[8]=xyt_handle->info->msg_buf_sz;
175b1c944f5SJed Brown   PCTFS_giop(vals,work,sizeof(op)/sizeof(op[0])-1,op);
176827bd09bSSatish Balay 
1772fa5cd67SKarl Rupp   fvals[0]=fvals[1]=fvals[2]=xyt_handle->info->tot_solve_time/xyt_handle->info->nsolves++;
178b1c944f5SJed Brown   PCTFS_grop(fvals,fwork,sizeof(fop)/sizeof(fop[0])-1,fop);
179827bd09bSSatish Balay 
180b1c944f5SJed Brown   if (!PCTFS_my_id) {
181b1c944f5SJed Brown     PetscPrintf(PETSC_COMM_WORLD,"%D :: min   xyt_nnz=%D\n",PCTFS_my_id,vals[0]);
182b1c944f5SJed Brown     PetscPrintf(PETSC_COMM_WORLD,"%D :: max   xyt_nnz=%D\n",PCTFS_my_id,vals[1]);
183b1c944f5SJed Brown     PetscPrintf(PETSC_COMM_WORLD,"%D :: avg   xyt_nnz=%g\n",PCTFS_my_id,1.0*vals[2]/PCTFS_num_nodes);
184b1c944f5SJed Brown     PetscPrintf(PETSC_COMM_WORLD,"%D :: tot   xyt_nnz=%D\n",PCTFS_my_id,vals[2]);
18577b4d14cSPeter Brune     PetscPrintf(PETSC_COMM_WORLD,"%D :: xyt   C(2d)  =%g\n",PCTFS_my_id,vals[2]/(PetscPowReal(1.0*vals[5],1.5)));
18677b4d14cSPeter Brune     PetscPrintf(PETSC_COMM_WORLD,"%D :: xyt   C(3d)  =%g\n",PCTFS_my_id,vals[2]/(PetscPowReal(1.0*vals[5],1.6667)));
187b1c944f5SJed Brown     PetscPrintf(PETSC_COMM_WORLD,"%D :: min   xyt_n  =%D\n",PCTFS_my_id,vals[3]);
188b1c944f5SJed Brown     PetscPrintf(PETSC_COMM_WORLD,"%D :: max   xyt_n  =%D\n",PCTFS_my_id,vals[4]);
189b1c944f5SJed Brown     PetscPrintf(PETSC_COMM_WORLD,"%D :: avg   xyt_n  =%g\n",PCTFS_my_id,1.0*vals[5]/PCTFS_num_nodes);
190b1c944f5SJed Brown     PetscPrintf(PETSC_COMM_WORLD,"%D :: tot   xyt_n  =%D\n",PCTFS_my_id,vals[5]);
191b1c944f5SJed Brown     PetscPrintf(PETSC_COMM_WORLD,"%D :: min   xyt_buf=%D\n",PCTFS_my_id,vals[6]);
192b1c944f5SJed Brown     PetscPrintf(PETSC_COMM_WORLD,"%D :: max   xyt_buf=%D\n",PCTFS_my_id,vals[7]);
193b1c944f5SJed Brown     PetscPrintf(PETSC_COMM_WORLD,"%D :: avg   xyt_buf=%g\n",PCTFS_my_id,1.0*vals[8]/PCTFS_num_nodes);
194b1c944f5SJed Brown     PetscPrintf(PETSC_COMM_WORLD,"%D :: min   xyt_slv=%g\n",PCTFS_my_id,fvals[0]);
195b1c944f5SJed Brown     PetscPrintf(PETSC_COMM_WORLD,"%D :: max   xyt_slv=%g\n",PCTFS_my_id,fvals[1]);
196b1c944f5SJed Brown     PetscPrintf(PETSC_COMM_WORLD,"%D :: avg   xyt_slv=%g\n",PCTFS_my_id,fvals[2]/PCTFS_num_nodes);
197827bd09bSSatish Balay   }
198827bd09bSSatish Balay 
199827bd09bSSatish Balay   return(0);
200827bd09bSSatish Balay }
201827bd09bSSatish Balay 
202827bd09bSSatish Balay /*************************************xyt.c************************************
203827bd09bSSatish Balay 
204827bd09bSSatish Balay Description: get A_local, local portion of global coarse matrix which
205827bd09bSSatish Balay is a row dist. nxm matrix w/ n<m.
206827bd09bSSatish Balay    o my_ml holds address of ML struct associated w/A_local and coarse grid
207827bd09bSSatish Balay    o local2global holds global number of column i (i=0,...,m-1)
208827bd09bSSatish Balay    o local2global holds global number of row    i (i=0,...,n-1)
209827bd09bSSatish Balay    o mylocmatvec performs A_local . vec_local (note that gs is performed using
210ca8e9878SJed Brown    PCTFS_gs_init/gop).
211827bd09bSSatish Balay 
212827bd09bSSatish Balay mylocmatvec = my_ml->Amat[grid_tag].matvec->external;
213827bd09bSSatish Balay mylocmatvec (void :: void *data, double *in, double *out)
214827bd09bSSatish Balay **************************************xyt.c***********************************/
21538dcbb09SBarry Smith static PetscErrorCode do_xyt_factor(xyt_ADT xyt_handle)
216827bd09bSSatish Balay {
2177b1ae94cSBarry Smith   return xyt_generate(xyt_handle);
218827bd09bSSatish Balay }
219827bd09bSSatish Balay 
2207b1ae94cSBarry Smith /**************************************xyt.c***********************************/
22138dcbb09SBarry Smith static PetscErrorCode xyt_generate(xyt_ADT xyt_handle)
222827bd09bSSatish Balay {
22352f87cdaSBarry Smith   PetscInt       i,j,k,idx;
22452f87cdaSBarry Smith   PetscInt       dim, col;
225a501084fSBarry Smith   PetscScalar    *u, *uu, *v, *z, *w, alpha, alpha_w;
22652f87cdaSBarry Smith   PetscInt       *segs;
22752f87cdaSBarry Smith   PetscInt       op[] = {GL_ADD,0};
22852f87cdaSBarry Smith   PetscInt       off, len;
229a501084fSBarry Smith   PetscScalar    *x_ptr, *y_ptr;
23052f87cdaSBarry Smith   PetscInt       *iptr, flag;
23152f87cdaSBarry Smith   PetscInt       start =0, end, work;
23252f87cdaSBarry Smith   PetscInt       op2[] = {GL_MIN,0};
233ca8e9878SJed Brown   PCTFS_gs_ADT   PCTFS_gs_handle;
23452f87cdaSBarry Smith   PetscInt       *nsep, *lnsep, *fo;
23552f87cdaSBarry Smith   PetscInt       a_n            =xyt_handle->mvi->n;
23652f87cdaSBarry Smith   PetscInt       a_m            =xyt_handle->mvi->m;
23752f87cdaSBarry Smith   PetscInt       *a_local2global=xyt_handle->mvi->local2global;
23852f87cdaSBarry Smith   PetscInt       level;
23952f87cdaSBarry Smith   PetscInt       n, m;
24052f87cdaSBarry Smith   PetscInt       *xcol_sz, *xcol_indices, *stages;
241a501084fSBarry Smith   PetscScalar    **xcol_vals, *x;
24252f87cdaSBarry Smith   PetscInt       *ycol_sz, *ycol_indices;
243a501084fSBarry Smith   PetscScalar    **ycol_vals, *y;
24452f87cdaSBarry Smith   PetscInt       n_global;
24552f87cdaSBarry Smith   PetscInt       xt_nnz       =0, xt_max_nnz=0;
24652f87cdaSBarry Smith   PetscInt       yt_nnz       =0, yt_max_nnz=0;
24752f87cdaSBarry Smith   PetscInt       xt_zero_nnz  =0;
24852f87cdaSBarry Smith   PetscInt       xt_zero_nnz_0=0;
24952f87cdaSBarry Smith   PetscInt       yt_zero_nnz  =0;
25052f87cdaSBarry Smith   PetscInt       yt_zero_nnz_0=0;
2514a0de3f6SBarry Smith   PetscBLASInt   i1           = 1,dlen;
252a501084fSBarry Smith   PetscScalar    dm1          = -1.0;
253d1528f56SBarry Smith   PetscErrorCode ierr;
254827bd09bSSatish Balay 
255827bd09bSSatish Balay   n              =xyt_handle->mvi->n;
256827bd09bSSatish Balay   nsep           =xyt_handle->info->nsep;
257827bd09bSSatish Balay   lnsep          =xyt_handle->info->lnsep;
258827bd09bSSatish Balay   fo             =xyt_handle->info->fo;
259827bd09bSSatish Balay   end            =lnsep[0];
260827bd09bSSatish Balay   level          =xyt_handle->level;
261ca8e9878SJed Brown   PCTFS_gs_handle=xyt_handle->mvi->PCTFS_gs_handle;
262827bd09bSSatish Balay 
263827bd09bSSatish Balay   /* is there a null space? */
264827bd09bSSatish Balay   /* LATER add in ability to detect null space by checking alpha */
2652fa5cd67SKarl Rupp   for (i=0, j=0; i<=level; i++) j+=nsep[i];
266827bd09bSSatish Balay 
267827bd09bSSatish Balay   m = j-xyt_handle->ns;
26822d28d08SBarry Smith   if (m!=j) {
26922d28d08SBarry Smith     ierr = PetscPrintf(PETSC_COMM_WORLD,"xyt_generate() :: null space exists %D %D %D\n",m,j,xyt_handle->ns);CHKERRQ(ierr);
27022d28d08SBarry Smith   }
271827bd09bSSatish Balay 
2727d3de750SJacob Faibussowitsch   ierr = PetscInfo(0,"xyt_generate() :: X(%D,%D)\n",n,m);CHKERRQ(ierr);
273827bd09bSSatish Balay 
274827bd09bSSatish Balay   /* get and initialize storage for x local         */
275827bd09bSSatish Balay   /* note that x local is nxm and stored by columns */
27652f87cdaSBarry Smith   xcol_sz      = (PetscInt*) malloc(m*sizeof(PetscInt));
27752f87cdaSBarry Smith   xcol_indices = (PetscInt*) malloc((2*m+1)*sizeof(PetscInt));
278a501084fSBarry Smith   xcol_vals    = (PetscScalar**) malloc(m*sizeof(PetscScalar*));
279db4deed7SKarl Rupp   for (i=j=0; i<m; i++, j+=2) {
280827bd09bSSatish Balay     xcol_indices[j]=xcol_indices[j+1]=xcol_sz[i]=-1;
281827bd09bSSatish Balay     xcol_vals[i]   = NULL;
282827bd09bSSatish Balay   }
283827bd09bSSatish Balay   xcol_indices[j]=-1;
284827bd09bSSatish Balay 
285827bd09bSSatish Balay   /* get and initialize storage for y local         */
286827bd09bSSatish Balay   /* note that y local is nxm and stored by columns */
28752f87cdaSBarry Smith   ycol_sz      = (PetscInt*) malloc(m*sizeof(PetscInt));
28852f87cdaSBarry Smith   ycol_indices = (PetscInt*) malloc((2*m+1)*sizeof(PetscInt));
289a501084fSBarry Smith   ycol_vals    = (PetscScalar**) malloc(m*sizeof(PetscScalar*));
290db4deed7SKarl Rupp   for (i=j=0; i<m; i++, j+=2) {
291827bd09bSSatish Balay     ycol_indices[j]=ycol_indices[j+1]=ycol_sz[i]=-1;
292827bd09bSSatish Balay     ycol_vals[i]   = NULL;
293827bd09bSSatish Balay   }
294827bd09bSSatish Balay   ycol_indices[j]=-1;
295827bd09bSSatish Balay 
296827bd09bSSatish Balay   /* size of separators for each sub-hc working from bottom of tree to top */
297827bd09bSSatish Balay   /* this looks like nsep[]=segments */
29852f87cdaSBarry Smith   stages = (PetscInt*) malloc((level+1)*sizeof(PetscInt));
29952f87cdaSBarry Smith   segs   = (PetscInt*) malloc((level+1)*sizeof(PetscInt));
300ca8e9878SJed Brown   PCTFS_ivec_zero(stages,level+1);
301ca8e9878SJed Brown   PCTFS_ivec_copy(segs,nsep,level+1);
3022fa5cd67SKarl Rupp   for (i=0; i<level; i++) segs[i+1] += segs[i];
303827bd09bSSatish Balay   stages[0] = segs[0];
304827bd09bSSatish Balay 
305827bd09bSSatish Balay   /* temporary vectors  */
306a501084fSBarry Smith   u  = (PetscScalar*) malloc(n*sizeof(PetscScalar));
307a501084fSBarry Smith   z  = (PetscScalar*) malloc(n*sizeof(PetscScalar));
308a501084fSBarry Smith   v  = (PetscScalar*) malloc(a_m*sizeof(PetscScalar));
309a501084fSBarry Smith   uu = (PetscScalar*) malloc(m*sizeof(PetscScalar));
310a501084fSBarry Smith   w  = (PetscScalar*) malloc(m*sizeof(PetscScalar));
311827bd09bSSatish Balay 
312827bd09bSSatish Balay   /* extra nnz due to replication of vertices across separators */
3132fa5cd67SKarl Rupp   for (i=1, j=0; i<=level; i++) j+=nsep[i];
314827bd09bSSatish Balay 
315827bd09bSSatish Balay   /* storage for sparse x values */
316827bd09bSSatish Balay   n_global   = xyt_handle->info->n_global;
31785ec1a3cSBarry Smith   xt_max_nnz = yt_max_nnz = (PetscInt)(2.5*PetscPowReal(1.0*n_global,1.6667) + j*n/2)/PCTFS_num_nodes;
318a501084fSBarry Smith   x          = (PetscScalar*) malloc(xt_max_nnz*sizeof(PetscScalar));
319a501084fSBarry Smith   y          = (PetscScalar*) malloc(yt_max_nnz*sizeof(PetscScalar));
320827bd09bSSatish Balay 
321827bd09bSSatish Balay   /* LATER - can embed next sep to fire in gs */
322827bd09bSSatish Balay   /* time to make the donuts - generate X factor */
323db4deed7SKarl Rupp   for (dim=i=j=0; i<m; i++) {
324827bd09bSSatish Balay     /* time to move to the next level? */
325d4af0d30SBarry Smith     while (i==segs[dim]) {
326*2c71b3e2SJacob Faibussowitsch       PetscCheckFalse(dim==level,PETSC_COMM_SELF,PETSC_ERR_PLIB,"dim about to exceed level");
327827bd09bSSatish Balay       stages[dim++]=i;
328827bd09bSSatish Balay       end         +=lnsep[dim];
329827bd09bSSatish Balay     }
330827bd09bSSatish Balay     stages[dim]=i;
331827bd09bSSatish Balay 
332827bd09bSSatish Balay     /* which column are we firing? */
333827bd09bSSatish Balay     /* i.e. set v_l */
334827bd09bSSatish Balay     /* use new seps and do global min across hc to determine which one to fire */
335827bd09bSSatish Balay     (start<end) ? (col=fo[start]) : (col=INT_MAX);
336b1c944f5SJed Brown     PCTFS_giop_hc(&col,&work,1,op2,dim);
337827bd09bSSatish Balay 
338827bd09bSSatish Balay     /* shouldn't need this */
339db4deed7SKarl Rupp     if (col==INT_MAX) {
340f1ed62a8SBarry Smith       ierr = PetscInfo(0,"hey ... col==INT_MAX??\n");CHKERRQ(ierr);
341827bd09bSSatish Balay       continue;
342827bd09bSSatish Balay     }
343827bd09bSSatish Balay 
344827bd09bSSatish Balay     /* do I own it? I should */
345ca8e9878SJed Brown     PCTFS_rvec_zero(v,a_m);
3462fa5cd67SKarl Rupp     if (col==fo[start]) {
347827bd09bSSatish Balay       start++;
348ca8e9878SJed Brown       idx=PCTFS_ivec_linear_search(col, a_local2global, a_n);
3492fa5cd67SKarl Rupp       if (idx!=-1) {
3502fa5cd67SKarl Rupp         v[idx] = 1.0;
3512fa5cd67SKarl Rupp         j++;
352546078acSJacob Faibussowitsch       } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"NOT FOUND!");
353db4deed7SKarl Rupp     } else {
354ca8e9878SJed Brown       idx=PCTFS_ivec_linear_search(col, a_local2global, a_m);
3552fa5cd67SKarl Rupp       if (idx!=-1) v[idx] = 1.0;
356827bd09bSSatish Balay     }
357827bd09bSSatish Balay 
358827bd09bSSatish Balay     /* perform u = A.v_l */
359ca8e9878SJed Brown     PCTFS_rvec_zero(u,n);
360827bd09bSSatish Balay     do_matvec(xyt_handle->mvi,v,u);
361827bd09bSSatish Balay 
362827bd09bSSatish Balay     /* uu =  X^T.u_l (local portion) */
363827bd09bSSatish Balay     /* technically only need to zero out first i entries */
364827bd09bSSatish Balay     /* later turn this into an XYT_solve call ? */
365ca8e9878SJed Brown     PCTFS_rvec_zero(uu,m);
366827bd09bSSatish Balay     y_ptr=y;
367827bd09bSSatish Balay     iptr = ycol_indices;
368db4deed7SKarl Rupp     for (k=0; k<i; k++) {
369827bd09bSSatish Balay       off   = *iptr++;
370827bd09bSSatish Balay       len   = *iptr++;
371c5df96a5SBarry Smith       ierr  = PetscBLASIntCast(len,&dlen);CHKERRQ(ierr);
3728b83055fSJed Brown       PetscStackCallBLAS("BLASdot",uu[k] = BLASdot_(&dlen,u+off,&i1,y_ptr,&i1));
373827bd09bSSatish Balay       y_ptr+=len;
374827bd09bSSatish Balay     }
375827bd09bSSatish Balay 
376827bd09bSSatish Balay     /* uu = X^T.u_l (comm portion) */
377a00740a5SBarry Smith     ierr = PCTFS_ssgl_radd  (uu, w, dim, stages);CHKERRQ(ierr);
378827bd09bSSatish Balay 
379827bd09bSSatish Balay     /* z = X.uu */
380ca8e9878SJed Brown     PCTFS_rvec_zero(z,n);
381827bd09bSSatish Balay     x_ptr=x;
382827bd09bSSatish Balay     iptr = xcol_indices;
383db4deed7SKarl Rupp     for (k=0; k<i; k++) {
384827bd09bSSatish Balay       off  = *iptr++;
385827bd09bSSatish Balay       len  = *iptr++;
386c5df96a5SBarry Smith       ierr = PetscBLASIntCast(len,&dlen);CHKERRQ(ierr);
3878b83055fSJed Brown       PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&dlen,&uu[k],x_ptr,&i1,z+off,&i1));
388827bd09bSSatish Balay       x_ptr+=len;
389827bd09bSSatish Balay     }
390827bd09bSSatish Balay 
391827bd09bSSatish Balay     /* compute v_l = v_l - z */
392ca8e9878SJed Brown     PCTFS_rvec_zero(v+a_n,a_m-a_n);
393c5df96a5SBarry Smith     ierr = PetscBLASIntCast(n,&dlen);CHKERRQ(ierr);
3948b83055fSJed Brown     PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&dlen,&dm1,z,&i1,v,&i1));
395827bd09bSSatish Balay 
396827bd09bSSatish Balay     /* compute u_l = A.v_l */
3972fa5cd67SKarl Rupp     if (a_n!=a_m) PCTFS_gs_gop_hc(PCTFS_gs_handle,v,"+\0",dim);
398ca8e9878SJed Brown     PCTFS_rvec_zero(u,n);
399827bd09bSSatish Balay     do_matvec(xyt_handle->mvi,v,u);
400827bd09bSSatish Balay 
401827bd09bSSatish Balay     /* compute sqrt(alpha) = sqrt(u_l^T.u_l) - local portion */
402c5df96a5SBarry Smith     ierr  = PetscBLASIntCast(n,&dlen);CHKERRQ(ierr);
4038b83055fSJed Brown     PetscStackCallBLAS("BLASdot",alpha = BLASdot_(&dlen,u,&i1,u,&i1));
404827bd09bSSatish Balay     /* compute sqrt(alpha) = sqrt(u_l^T.u_l) - comm portion */
405b1c944f5SJed Brown     PCTFS_grop_hc(&alpha, &alpha_w, 1, op, dim);
406827bd09bSSatish Balay 
4078f1a2a5eSBarry Smith     alpha = (PetscScalar) PetscSqrtReal((PetscReal)alpha);
408827bd09bSSatish Balay 
409827bd09bSSatish Balay     /* check for small alpha                             */
410827bd09bSSatish Balay     /* LATER use this to detect and determine null space */
411*2c71b3e2SJacob Faibussowitsch     PetscCheckFalse(PetscAbsScalar(alpha)<1.0e-14,PETSC_COMM_SELF,PETSC_ERR_PLIB,"bad alpha! %g",alpha);
412827bd09bSSatish Balay 
413827bd09bSSatish Balay     /* compute v_l = v_l/sqrt(alpha) */
414ca8e9878SJed Brown     PCTFS_rvec_scale(v,1.0/alpha,n);
415ca8e9878SJed Brown     PCTFS_rvec_scale(u,1.0/alpha,n);
416827bd09bSSatish Balay 
417827bd09bSSatish Balay     /* add newly generated column, v_l, to X */
418827bd09bSSatish Balay     flag = 1;
419827bd09bSSatish Balay     off  =len=0;
420db4deed7SKarl Rupp     for (k=0; k<n; k++) {
421db4deed7SKarl Rupp       if (v[k]!=0.0) {
422827bd09bSSatish Balay         len=k;
423db4deed7SKarl Rupp         if (flag) {off=k; flag=0;}
424827bd09bSSatish Balay       }
425827bd09bSSatish Balay     }
426827bd09bSSatish Balay 
427827bd09bSSatish Balay     len -= (off-1);
428827bd09bSSatish Balay 
429db4deed7SKarl Rupp     if (len>0) {
430db4deed7SKarl Rupp       if ((xt_nnz+len)>xt_max_nnz) {
431f1ed62a8SBarry Smith         ierr        = PetscInfo(0,"increasing space for X by 2x!\n");CHKERRQ(ierr);
432827bd09bSSatish Balay         xt_max_nnz *= 2;
433a501084fSBarry Smith         x_ptr       = (PetscScalar*) malloc(xt_max_nnz*sizeof(PetscScalar));
434ca8e9878SJed Brown         PCTFS_rvec_copy(x_ptr,x,xt_nnz);
435a501084fSBarry Smith         free(x);
436827bd09bSSatish Balay         x     = x_ptr;
437827bd09bSSatish Balay         x_ptr+=xt_nnz;
438827bd09bSSatish Balay       }
439827bd09bSSatish Balay       xt_nnz += len;
440ca8e9878SJed Brown       PCTFS_rvec_copy(x_ptr,v+off,len);
441827bd09bSSatish Balay 
442827bd09bSSatish Balay       /* keep track of number of zeros */
443db4deed7SKarl Rupp       if (dim) {
444db4deed7SKarl Rupp         for (k=0; k<len; k++) {
4452fa5cd67SKarl Rupp           if (x_ptr[k]==0.0) xt_zero_nnz++;
446827bd09bSSatish Balay         }
447db4deed7SKarl Rupp       } else {
448db4deed7SKarl Rupp         for (k=0; k<len; k++) {
4492fa5cd67SKarl Rupp           if (x_ptr[k]==0.0) xt_zero_nnz_0++;
450827bd09bSSatish Balay         }
451827bd09bSSatish Balay       }
452827bd09bSSatish Balay       xcol_indices[2*i] = off;
453827bd09bSSatish Balay       xcol_sz[i]        = xcol_indices[2*i+1] = len;
454827bd09bSSatish Balay       xcol_vals[i]      = x_ptr;
455db4deed7SKarl Rupp     } else {
456827bd09bSSatish Balay       xcol_indices[2*i] = 0;
457827bd09bSSatish Balay       xcol_sz[i]        = xcol_indices[2*i+1] = 0;
458827bd09bSSatish Balay       xcol_vals[i]      = x_ptr;
459827bd09bSSatish Balay     }
460827bd09bSSatish Balay 
461827bd09bSSatish Balay     /* add newly generated column, u_l, to Y */
462827bd09bSSatish Balay     flag = 1;
463827bd09bSSatish Balay     off  =len=0;
464db4deed7SKarl Rupp     for (k=0; k<n; k++) {
465db4deed7SKarl Rupp       if (u[k]!=0.0) {
466827bd09bSSatish Balay         len=k;
467db4deed7SKarl Rupp         if (flag) { off=k; flag=0; }
468827bd09bSSatish Balay       }
469827bd09bSSatish Balay     }
470827bd09bSSatish Balay 
471827bd09bSSatish Balay     len -= (off-1);
472827bd09bSSatish Balay 
473db4deed7SKarl Rupp     if (len>0) {
474db4deed7SKarl Rupp       if ((yt_nnz+len)>yt_max_nnz) {
475f1ed62a8SBarry Smith         ierr        = PetscInfo(0,"increasing space for Y by 2x!\n");CHKERRQ(ierr);
476827bd09bSSatish Balay         yt_max_nnz *= 2;
477a501084fSBarry Smith         y_ptr       = (PetscScalar*) malloc(yt_max_nnz*sizeof(PetscScalar));
478ca8e9878SJed Brown         PCTFS_rvec_copy(y_ptr,y,yt_nnz);
479a501084fSBarry Smith         free(y);
480827bd09bSSatish Balay         y     = y_ptr;
481827bd09bSSatish Balay         y_ptr+=yt_nnz;
482827bd09bSSatish Balay       }
483827bd09bSSatish Balay       yt_nnz += len;
484ca8e9878SJed Brown       PCTFS_rvec_copy(y_ptr,u+off,len);
485827bd09bSSatish Balay 
486827bd09bSSatish Balay       /* keep track of number of zeros */
487db4deed7SKarl Rupp       if (dim) {
488db4deed7SKarl Rupp         for (k=0; k<len; k++) {
4892fa5cd67SKarl Rupp           if (y_ptr[k]==0.0) yt_zero_nnz++;
490827bd09bSSatish Balay         }
491db4deed7SKarl Rupp       } else {
492db4deed7SKarl Rupp         for (k=0; k<len; k++) {
4932fa5cd67SKarl Rupp           if (y_ptr[k]==0.0) yt_zero_nnz_0++;
494827bd09bSSatish Balay         }
495827bd09bSSatish Balay       }
496827bd09bSSatish Balay       ycol_indices[2*i] = off;
497827bd09bSSatish Balay       ycol_sz[i]        = ycol_indices[2*i+1] = len;
498827bd09bSSatish Balay       ycol_vals[i]      = y_ptr;
499db4deed7SKarl Rupp     } else {
500827bd09bSSatish Balay       ycol_indices[2*i] = 0;
501827bd09bSSatish Balay       ycol_sz[i]        = ycol_indices[2*i+1] = 0;
502827bd09bSSatish Balay       ycol_vals[i]      = y_ptr;
503827bd09bSSatish Balay     }
504827bd09bSSatish Balay   }
505827bd09bSSatish Balay 
506827bd09bSSatish Balay   /* close off stages for execution phase */
507db4deed7SKarl Rupp   while (dim!=level) {
508827bd09bSSatish Balay     stages[dim++]=i;
5097d3de750SJacob Faibussowitsch     ierr         = PetscInfo(0,"disconnected!!! dim(%D)!=level(%D)\n",dim,level);CHKERRQ(ierr);
510827bd09bSSatish Balay   }
511827bd09bSSatish Balay   stages[dim]=i;
512827bd09bSSatish Balay 
513827bd09bSSatish Balay   xyt_handle->info->n           =xyt_handle->mvi->n;
514827bd09bSSatish Balay   xyt_handle->info->m           =m;
515827bd09bSSatish Balay   xyt_handle->info->nnz         =xt_nnz + yt_nnz;
516827bd09bSSatish Balay   xyt_handle->info->max_nnz     =xt_max_nnz + yt_max_nnz;
517827bd09bSSatish Balay   xyt_handle->info->msg_buf_sz  =stages[level]-stages[0];
518a501084fSBarry Smith   xyt_handle->info->solve_uu    = (PetscScalar*) malloc(m*sizeof(PetscScalar));
519a501084fSBarry Smith   xyt_handle->info->solve_w     = (PetscScalar*) malloc(m*sizeof(PetscScalar));
520827bd09bSSatish Balay   xyt_handle->info->x           =x;
521827bd09bSSatish Balay   xyt_handle->info->xcol_vals   =xcol_vals;
522827bd09bSSatish Balay   xyt_handle->info->xcol_sz     =xcol_sz;
523827bd09bSSatish Balay   xyt_handle->info->xcol_indices=xcol_indices;
524827bd09bSSatish Balay   xyt_handle->info->stages      =stages;
525827bd09bSSatish Balay   xyt_handle->info->y           =y;
526827bd09bSSatish Balay   xyt_handle->info->ycol_vals   =ycol_vals;
527827bd09bSSatish Balay   xyt_handle->info->ycol_sz     =ycol_sz;
528827bd09bSSatish Balay   xyt_handle->info->ycol_indices=ycol_indices;
529827bd09bSSatish Balay 
530a501084fSBarry Smith   free(segs);
531a501084fSBarry Smith   free(u);
532a501084fSBarry Smith   free(v);
533a501084fSBarry Smith   free(uu);
534a501084fSBarry Smith   free(z);
535a501084fSBarry Smith   free(w);
536827bd09bSSatish Balay 
537827bd09bSSatish Balay   return(0);
538827bd09bSSatish Balay }
539827bd09bSSatish Balay 
5407b1ae94cSBarry Smith /**************************************xyt.c***********************************/
5410924e98cSBarry Smith static PetscErrorCode do_xyt_solve(xyt_ADT xyt_handle,  PetscScalar *uc)
542827bd09bSSatish Balay {
54352f87cdaSBarry Smith   PetscInt       off, len, *iptr;
54452f87cdaSBarry Smith   PetscInt       level        =xyt_handle->level;
54552f87cdaSBarry Smith   PetscInt       n            =xyt_handle->info->n;
54652f87cdaSBarry Smith   PetscInt       m            =xyt_handle->info->m;
54752f87cdaSBarry Smith   PetscInt       *stages      =xyt_handle->info->stages;
54852f87cdaSBarry Smith   PetscInt       *xcol_indices=xyt_handle->info->xcol_indices;
54952f87cdaSBarry Smith   PetscInt       *ycol_indices=xyt_handle->info->ycol_indices;
550a501084fSBarry Smith   PetscScalar    *x_ptr, *y_ptr, *uu_ptr;
551a501084fSBarry Smith   PetscScalar    *solve_uu=xyt_handle->info->solve_uu;
552a501084fSBarry Smith   PetscScalar    *solve_w =xyt_handle->info->solve_w;
553a501084fSBarry Smith   PetscScalar    *x       =xyt_handle->info->x;
554a501084fSBarry Smith   PetscScalar    *y       =xyt_handle->info->y;
5554a0de3f6SBarry Smith   PetscBLASInt   i1       = 1,dlen;
556c5df96a5SBarry Smith   PetscErrorCode ierr;
557827bd09bSSatish Balay 
5580924e98cSBarry Smith   PetscFunctionBegin;
559827bd09bSSatish Balay   uu_ptr=solve_uu;
560ca8e9878SJed Brown   PCTFS_rvec_zero(uu_ptr,m);
561827bd09bSSatish Balay 
562827bd09bSSatish Balay   /* x  = X.Y^T.b */
563827bd09bSSatish Balay   /* uu = Y^T.b */
564947b95d8SBarry Smith   for (y_ptr=y,iptr=ycol_indices; *iptr!=-1; y_ptr+=len) {
565c5df96a5SBarry Smith     off       =*iptr++;
566c5df96a5SBarry Smith     len       =*iptr++;
567c5df96a5SBarry Smith     ierr      = PetscBLASIntCast(len,&dlen);CHKERRQ(ierr);
5688b83055fSJed Brown     PetscStackCallBLAS("BLASdot",*uu_ptr++ = BLASdot_(&dlen,uc+off,&i1,y_ptr,&i1));
569827bd09bSSatish Balay   }
570827bd09bSSatish Balay 
571827bd09bSSatish Balay   /* comunication of beta */
572827bd09bSSatish Balay   uu_ptr=solve_uu;
573a00740a5SBarry Smith   if (level) {ierr = PCTFS_ssgl_radd(uu_ptr, solve_w, level, stages);CHKERRQ(ierr);}
574ca8e9878SJed Brown   PCTFS_rvec_zero(uc,n);
575827bd09bSSatish Balay 
576827bd09bSSatish Balay   /* x = X.uu */
577db4deed7SKarl Rupp   for (x_ptr=x,iptr=xcol_indices; *iptr!=-1; x_ptr+=len) {
578c5df96a5SBarry Smith     off  =*iptr++;
579c5df96a5SBarry Smith     len  =*iptr++;
580c5df96a5SBarry Smith     ierr = PetscBLASIntCast(len,&dlen);CHKERRQ(ierr);
5818b83055fSJed Brown     PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&dlen,uu_ptr++,x_ptr,&i1,uc+off,&i1));
582827bd09bSSatish Balay   }
5830924e98cSBarry Smith   PetscFunctionReturn(0);
584827bd09bSSatish Balay }
585827bd09bSSatish Balay 
5867b1ae94cSBarry Smith /**************************************xyt.c***********************************/
5870924e98cSBarry Smith static PetscErrorCode check_handle(xyt_ADT xyt_handle)
588827bd09bSSatish Balay {
58952f87cdaSBarry Smith   PetscInt vals[2], work[2], op[] = {NON_UNIFORM,GL_MIN,GL_MAX};
590827bd09bSSatish Balay 
5910924e98cSBarry Smith   PetscFunctionBegin;
592*2c71b3e2SJacob Faibussowitsch   PetscCheckFalse(!xyt_handle,PETSC_COMM_SELF,PETSC_ERR_PLIB,"check_handle() :: bad handle :: NULL %D",xyt_handle);
593827bd09bSSatish Balay 
594827bd09bSSatish Balay   vals[0]=vals[1]=xyt_handle->id;
595b1c944f5SJed Brown   PCTFS_giop(vals,work,sizeof(op)/sizeof(op[0])-1,op);
596*2c71b3e2SJacob Faibussowitsch   PetscCheckFalse((vals[0]!=vals[1])||(xyt_handle->id<=0),PETSC_COMM_SELF,PETSC_ERR_PLIB,"check_handle() :: bad handle :: id mismatch min/max %D/%D %D", vals[0],vals[1], xyt_handle->id);
5970924e98cSBarry Smith   PetscFunctionReturn(0);
598827bd09bSSatish Balay }
599827bd09bSSatish Balay 
6007b1ae94cSBarry Smith /**************************************xyt.c***********************************/
6010924e98cSBarry Smith static PetscErrorCode det_separators(xyt_ADT xyt_handle)
602827bd09bSSatish Balay {
60352f87cdaSBarry Smith   PetscInt       i, ct, id;
60452f87cdaSBarry Smith   PetscInt       mask, edge, *iptr;
60552f87cdaSBarry Smith   PetscInt       *dir, *used;
60652f87cdaSBarry Smith   PetscInt       sum[4], w[4];
607a501084fSBarry Smith   PetscScalar    rsum[4], rw[4];
60852f87cdaSBarry Smith   PetscInt       op[] = {GL_ADD,0};
609a501084fSBarry Smith   PetscScalar    *lhs, *rhs;
61052f87cdaSBarry Smith   PetscInt       *nsep, *lnsep, *fo, nfo=0;
611ca8e9878SJed Brown   PCTFS_gs_ADT   PCTFS_gs_handle=xyt_handle->mvi->PCTFS_gs_handle;
61252f87cdaSBarry Smith   PetscInt       *local2global  =xyt_handle->mvi->local2global;
61352f87cdaSBarry Smith   PetscInt       n              =xyt_handle->mvi->n;
61452f87cdaSBarry Smith   PetscInt       m              =xyt_handle->mvi->m;
61552f87cdaSBarry Smith   PetscInt       level          =xyt_handle->level;
616ab824b78SBarry Smith   PetscInt       shared         =0;
617d1528f56SBarry Smith   PetscErrorCode ierr;
618827bd09bSSatish Balay 
6190924e98cSBarry Smith   PetscFunctionBegin;
62052f87cdaSBarry Smith   dir  = (PetscInt*)malloc(sizeof(PetscInt)*(level+1));
62152f87cdaSBarry Smith   nsep = (PetscInt*)malloc(sizeof(PetscInt)*(level+1));
62252f87cdaSBarry Smith   lnsep= (PetscInt*)malloc(sizeof(PetscInt)*(level+1));
62352f87cdaSBarry Smith   fo   = (PetscInt*)malloc(sizeof(PetscInt)*(n+1));
62452f87cdaSBarry Smith   used = (PetscInt*)malloc(sizeof(PetscInt)*n);
625827bd09bSSatish Balay 
626ca8e9878SJed Brown   PCTFS_ivec_zero(dir,level+1);
627ca8e9878SJed Brown   PCTFS_ivec_zero(nsep,level+1);
628ca8e9878SJed Brown   PCTFS_ivec_zero(lnsep,level+1);
629ca8e9878SJed Brown   PCTFS_ivec_set (fo,-1,n+1);
630ca8e9878SJed Brown   PCTFS_ivec_zero(used,n);
631827bd09bSSatish Balay 
6328cda6cd7SBarry Smith   lhs = (PetscScalar*)malloc(sizeof(PetscScalar)*m);
6338cda6cd7SBarry Smith   rhs = (PetscScalar*)malloc(sizeof(PetscScalar)*m);
634827bd09bSSatish Balay 
635827bd09bSSatish Balay   /* determine the # of unique dof */
636ca8e9878SJed Brown   PCTFS_rvec_zero(lhs,m);
637ca8e9878SJed Brown   PCTFS_rvec_set(lhs,1.0,n);
638ca8e9878SJed Brown   PCTFS_gs_gop_hc(PCTFS_gs_handle,lhs,"+\0",level);
639ca8e9878SJed Brown   ierr = PetscInfo(0,"done first PCTFS_gs_gop_hc\n");CHKERRQ(ierr);
640ca8e9878SJed Brown   PCTFS_rvec_zero(rsum,2);
641947b95d8SBarry Smith   for (i=0; i<n; i++) {
642db4deed7SKarl Rupp     if (lhs[i]!=0.0) { rsum[0]+=1.0/lhs[i]; rsum[1]+=lhs[i]; }
6432fa5cd67SKarl Rupp     if (lhs[i]!=1.0) shared=1;
644827bd09bSSatish Balay   }
645827bd09bSSatish Balay 
646b1c944f5SJed Brown   PCTFS_grop_hc(rsum,rw,2,op,level);
647827bd09bSSatish Balay   rsum[0]+=0.1;
648827bd09bSSatish Balay   rsum[1]+=0.1;
649827bd09bSSatish Balay 
65052f87cdaSBarry Smith   xyt_handle->info->n_global=xyt_handle->info->m_global=(PetscInt) rsum[0];
65152f87cdaSBarry Smith   xyt_handle->mvi->n_global =xyt_handle->mvi->m_global =(PetscInt) rsum[0];
652827bd09bSSatish Balay 
653827bd09bSSatish Balay   /* determine separator sets top down */
6542fa5cd67SKarl Rupp   if (shared) {
655827bd09bSSatish Balay     /* solution is to do as in the symmetric shared case but then */
656827bd09bSSatish Balay     /* pick the sub-hc with the most free dofs and do a mat-vec   */
657827bd09bSSatish Balay     /* and pick up the responses on the other sub-hc from the     */
658827bd09bSSatish Balay     /* initial separator set obtained from the symm. shared case  */
659546078acSJacob Faibussowitsch     SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"shared dof separator determination not ready ... see hmt!!!");
6604e619c20SJed Brown     /* [dead code deleted since it is unlikely to be completed] */
661db4deed7SKarl Rupp   } else {
662db4deed7SKarl Rupp     for (iptr=fo+n,id=PCTFS_my_id,mask=PCTFS_num_nodes>>1,edge=level;edge>0;edge--,mask>>=1) {
663827bd09bSSatish Balay       /* set rsh of hc, fire, and collect lhs responses */
664ca8e9878SJed Brown       (id<mask) ? PCTFS_rvec_zero(lhs,m) : PCTFS_rvec_set(lhs,1.0,m);
665ca8e9878SJed Brown       PCTFS_gs_gop_hc(PCTFS_gs_handle,lhs,"+\0",edge);
666827bd09bSSatish Balay 
667827bd09bSSatish Balay       /* set lsh of hc, fire, and collect rhs responses */
668ca8e9878SJed Brown       (id<mask) ? PCTFS_rvec_set(rhs,1.0,m) : PCTFS_rvec_zero(rhs,m);
669ca8e9878SJed Brown       PCTFS_gs_gop_hc(PCTFS_gs_handle,rhs,"+\0",edge);
670827bd09bSSatish Balay 
671827bd09bSSatish Balay       /* count number of dofs I own that have signal and not in sep set */
672db4deed7SKarl Rupp       for (PCTFS_ivec_zero(sum,4),ct=i=0;i<n;i++) {
673db4deed7SKarl Rupp         if (!used[i]) {
674827bd09bSSatish Balay           /* number of unmarked dofs on node */
675827bd09bSSatish Balay           ct++;
676827bd09bSSatish Balay           /* number of dofs to be marked on lhs hc */
6772fa5cd67SKarl Rupp           if ((id< mask)&&(lhs[i]!=0.0)) sum[0]++;
678827bd09bSSatish Balay           /* number of dofs to be marked on rhs hc */
6792fa5cd67SKarl Rupp           if ((id>=mask)&&(rhs[i]!=0.0)) sum[1]++;
680827bd09bSSatish Balay         }
681827bd09bSSatish Balay       }
682827bd09bSSatish Balay 
683827bd09bSSatish Balay       /* for the non-symmetric case we need separators of width 2 */
684827bd09bSSatish Balay       /* so take both sides */
685827bd09bSSatish Balay       (id<mask) ? (sum[2]=ct) : (sum[3]=ct);
686b1c944f5SJed Brown       PCTFS_giop_hc(sum,w,4,op,edge);
687827bd09bSSatish Balay 
688827bd09bSSatish Balay       ct=0;
689db4deed7SKarl Rupp       if (id<mask) {
690827bd09bSSatish Balay         /* mark dofs I own that have signal and not in sep set */
691db4deed7SKarl Rupp         for (i=0;i<n;i++) {
692db4deed7SKarl Rupp           if ((!used[i])&&(lhs[i]!=0.0)) {
693827bd09bSSatish Balay             ct++; nfo++;
694827bd09bSSatish Balay             *--iptr = local2global[i];
695827bd09bSSatish Balay             used[i] =edge;
696827bd09bSSatish Balay           }
697827bd09bSSatish Balay         }
698827bd09bSSatish Balay         /* LSH hc summation of ct should be sum[0] */
699db4deed7SKarl Rupp       } else {
700827bd09bSSatish Balay         /* mark dofs I own that have signal and not in sep set */
701db4deed7SKarl Rupp         for (i=0; i<n; i++) {
702db4deed7SKarl Rupp           if ((!used[i])&&(rhs[i]!=0.0)) {
703827bd09bSSatish Balay             ct++; nfo++;
704827bd09bSSatish Balay             *--iptr = local2global[i];
705827bd09bSSatish Balay             used[i] = edge;
706827bd09bSSatish Balay           }
707827bd09bSSatish Balay         }
708827bd09bSSatish Balay         /* RSH hc summation of ct should be sum[1] */
709827bd09bSSatish Balay       }
710827bd09bSSatish Balay 
7112fa5cd67SKarl Rupp       if (ct>1) PCTFS_ivec_sort(iptr,ct);
712827bd09bSSatish Balay       lnsep[edge]=ct;
713827bd09bSSatish Balay       nsep[edge] =sum[0]+sum[1];
714827bd09bSSatish Balay       dir [edge] =BOTH;
715827bd09bSSatish Balay 
716827bd09bSSatish Balay       /* LATER or we can recur on these to order seps at this level */
717827bd09bSSatish Balay       /* do we need full set of separators for this?                */
718827bd09bSSatish Balay 
719827bd09bSSatish Balay       /* fold rhs hc into lower */
7202fa5cd67SKarl Rupp       if (id>=mask) id-=mask;
721827bd09bSSatish Balay     }
722827bd09bSSatish Balay   }
723827bd09bSSatish Balay 
724827bd09bSSatish Balay   /* level 0 is on processor case - so mark the remainder */
725db4deed7SKarl Rupp   for (ct=i=0;i<n;i++) {
726db4deed7SKarl Rupp     if (!used[i]) {
727827bd09bSSatish Balay       ct++; nfo++;
728827bd09bSSatish Balay       *--iptr = local2global[i];
729827bd09bSSatish Balay       used[i] = edge;
730827bd09bSSatish Balay     }
731827bd09bSSatish Balay   }
7322fa5cd67SKarl Rupp   if (ct>1) PCTFS_ivec_sort(iptr,ct);
733827bd09bSSatish Balay   lnsep[edge]=ct;
734827bd09bSSatish Balay   nsep [edge]=ct;
735827bd09bSSatish Balay   dir  [edge]=BOTH;
736827bd09bSSatish Balay 
737827bd09bSSatish Balay   xyt_handle->info->nsep  = nsep;
738827bd09bSSatish Balay   xyt_handle->info->lnsep = lnsep;
739827bd09bSSatish Balay   xyt_handle->info->fo    = fo;
740827bd09bSSatish Balay   xyt_handle->info->nfo   = nfo;
741827bd09bSSatish Balay 
742a501084fSBarry Smith   free(dir);
743a501084fSBarry Smith   free(lhs);
744a501084fSBarry Smith   free(rhs);
745a501084fSBarry Smith   free(used);
7460924e98cSBarry Smith   PetscFunctionReturn(0);
747827bd09bSSatish Balay }
748827bd09bSSatish Balay 
7497b1ae94cSBarry Smith /**************************************xyt.c***********************************/
7505c8f6a95SKarl Rupp static mv_info *set_mvi(PetscInt *local2global, PetscInt n, PetscInt m, PetscErrorCode (*matvec)(mv_info*,PetscScalar*,PetscScalar*), void *grid_data)
751827bd09bSSatish Balay {
752827bd09bSSatish Balay   mv_info *mvi;
753827bd09bSSatish Balay 
754a501084fSBarry Smith   mvi              = (mv_info*)malloc(sizeof(mv_info));
755827bd09bSSatish Balay   mvi->n           = n;
756827bd09bSSatish Balay   mvi->m           = m;
757827bd09bSSatish Balay   mvi->n_global    = -1;
758827bd09bSSatish Balay   mvi->m_global    = -1;
75952f87cdaSBarry Smith   mvi->local2global= (PetscInt*)malloc((m+1)*sizeof(PetscInt));
7602fa5cd67SKarl Rupp 
761ca8e9878SJed Brown   PCTFS_ivec_copy(mvi->local2global,local2global,m);
762827bd09bSSatish Balay   mvi->local2global[m] = INT_MAX;
7635c8f6a95SKarl Rupp   mvi->matvec          = matvec;
764827bd09bSSatish Balay   mvi->grid_data       = grid_data;
765827bd09bSSatish Balay 
766827bd09bSSatish Balay   /* set xyt communication handle to perform restricted matvec */
767ca8e9878SJed Brown   mvi->PCTFS_gs_handle = PCTFS_gs_init(local2global, m, PCTFS_num_nodes);
768827bd09bSSatish Balay 
769827bd09bSSatish Balay   return(mvi);
770827bd09bSSatish Balay }
771827bd09bSSatish Balay 
7727b1ae94cSBarry Smith /**************************************xyt.c***********************************/
7733fdc5746SBarry Smith static PetscErrorCode do_matvec(mv_info *A, PetscScalar *v, PetscScalar *u)
774827bd09bSSatish Balay {
7750924e98cSBarry Smith   PetscFunctionBegin;
776827bd09bSSatish Balay   A->matvec((mv_info*)A->grid_data,v,u);
7770924e98cSBarry Smith   PetscFunctionReturn(0);
778827bd09bSSatish Balay }
779827bd09bSSatish Balay 
780