xref: /petsc/src/ksp/pc/impls/tfs/xxt.c (revision 5a586d828132823a7f656d0d7f79985f923612fa)
1dba47a55SKris Buschelman #define PETSCKSP_DLL
2827bd09bSSatish Balay 
3827bd09bSSatish Balay /*************************************xxt.c************************************
4827bd09bSSatish Balay Module Name: xxt
5827bd09bSSatish Balay Module Info:
6827bd09bSSatish Balay 
7827bd09bSSatish Balay author:  Henry M. Tufo III
8827bd09bSSatish Balay e-mail:  hmt@asci.uchicago.edu
9827bd09bSSatish Balay contact:
10827bd09bSSatish Balay +--------------------------------+--------------------------------+
11827bd09bSSatish Balay |MCS Division - Building 221     |Department of Computer Science  |
12827bd09bSSatish Balay |Argonne National Laboratory     |Ryerson 152                     |
13827bd09bSSatish Balay |9700 S. Cass Avenue             |The University of Chicago       |
14827bd09bSSatish Balay |Argonne, IL  60439              |Chicago, IL  60637              |
15827bd09bSSatish Balay |(630) 252-5354/5986 ph/fx       |(773) 702-6019/8487 ph/fx       |
16827bd09bSSatish Balay +--------------------------------+--------------------------------+
17827bd09bSSatish Balay 
18827bd09bSSatish Balay Last Modification: 3.20.01
19827bd09bSSatish Balay **************************************xxt.c***********************************/
207c4f633dSBarry Smith #include "../src/ksp/pc/impls/tfs/tfs.h"
21827bd09bSSatish Balay 
22827bd09bSSatish Balay #define LEFT  -1
23827bd09bSSatish Balay #define RIGHT  1
24827bd09bSSatish Balay #define BOTH   0
25827bd09bSSatish Balay 
26827bd09bSSatish Balay typedef struct xxt_solver_info {
2752f87cdaSBarry Smith   PetscInt n, m, n_global, m_global;
2852f87cdaSBarry Smith   PetscInt nnz, max_nnz, msg_buf_sz;
2952f87cdaSBarry Smith   PetscInt *nsep, *lnsep, *fo, nfo, *stages;
3052f87cdaSBarry Smith   PetscInt *col_sz, *col_indices;
31a501084fSBarry Smith   PetscScalar **col_vals, *x, *solve_uu, *solve_w;
3252f87cdaSBarry Smith   PetscInt nsolves;
33a501084fSBarry Smith   PetscScalar tot_solve_time;
34827bd09bSSatish Balay } xxt_info;
35827bd09bSSatish Balay 
36827bd09bSSatish Balay typedef struct matvec_info {
3752f87cdaSBarry Smith   PetscInt n, m, n_global, m_global;
3852f87cdaSBarry Smith   PetscInt *local2global;
39827bd09bSSatish Balay   gs_ADT gs_handle;
40a501084fSBarry Smith   PetscErrorCode (*matvec)(struct matvec_info*,PetscScalar*,PetscScalar*);
41827bd09bSSatish Balay   void *grid_data;
42827bd09bSSatish Balay } mv_info;
43827bd09bSSatish Balay 
44827bd09bSSatish Balay struct xxt_CDT{
4552f87cdaSBarry Smith   PetscInt id;
4652f87cdaSBarry Smith   PetscInt ns;
4752f87cdaSBarry Smith   PetscInt level;
48827bd09bSSatish Balay   xxt_info *info;
49827bd09bSSatish Balay   mv_info  *mvi;
50827bd09bSSatish Balay };
51827bd09bSSatish Balay 
5252f87cdaSBarry Smith static PetscInt n_xxt=0;
5352f87cdaSBarry Smith static PetscInt n_xxt_handles=0;
54827bd09bSSatish Balay 
55827bd09bSSatish Balay /* prototypes */
563fdc5746SBarry Smith static PetscErrorCode do_xxt_solve(xxt_ADT xxt_handle, PetscScalar *rhs);
573fdc5746SBarry Smith static PetscErrorCode check_handle(xxt_ADT xxt_handle);
583fdc5746SBarry Smith static PetscErrorCode det_separators(xxt_ADT xxt_handle);
593fdc5746SBarry Smith static PetscErrorCode do_matvec(mv_info *A, PetscScalar *v, PetscScalar *u);
6052f87cdaSBarry Smith static PetscInt xxt_generate(xxt_ADT xxt_handle);
6152f87cdaSBarry Smith static PetscInt do_xxt_factor(xxt_ADT xxt_handle);
6252f87cdaSBarry Smith static mv_info *set_mvi(PetscInt *local2global, PetscInt n, PetscInt m, void *matvec, void *grid_data);
63a501084fSBarry Smith 
647b1ae94cSBarry Smith /**************************************xxt.c***********************************/
657b1ae94cSBarry Smith xxt_ADT XXT_new(void)
66827bd09bSSatish Balay {
67827bd09bSSatish Balay   xxt_ADT xxt_handle;
68827bd09bSSatish Balay 
69827bd09bSSatish Balay   /* rolling count on n_xxt ... pot. problem here */
70827bd09bSSatish Balay   n_xxt_handles++;
71a501084fSBarry Smith   xxt_handle       = (xxt_ADT)malloc(sizeof(struct xxt_CDT));
72827bd09bSSatish Balay   xxt_handle->id   = ++n_xxt;
73827bd09bSSatish Balay   xxt_handle->info = NULL; xxt_handle->mvi  = NULL;
74827bd09bSSatish Balay 
75827bd09bSSatish Balay   return(xxt_handle);
76827bd09bSSatish Balay }
77827bd09bSSatish Balay 
787b1ae94cSBarry Smith /**************************************xxt.c***********************************/
7952f87cdaSBarry Smith PetscInt XXT_factor(xxt_ADT xxt_handle, /* prev. allocated xxt  handle */
8052f87cdaSBarry Smith 	   PetscInt *local2global,  /* global column mapping       */
8152f87cdaSBarry Smith 	   PetscInt n,              /* local num rows              */
8252f87cdaSBarry Smith 	   PetscInt m,              /* local num cols              */
83827bd09bSSatish Balay 	   void *matvec,       /* b_loc=A_local.x_loc         */
84827bd09bSSatish Balay 	   void *grid_data     /* grid data for matvec        */
85827bd09bSSatish Balay 	   )
86827bd09bSSatish Balay {
87a40341efSBarry Smith   comm_init();
88827bd09bSSatish Balay   check_handle(xxt_handle);
89827bd09bSSatish Balay 
90827bd09bSSatish Balay   /* only 2^k for now and all nodes participating */
91c1235816SBarry Smith   if ((1<<(xxt_handle->level=i_log2_num_nodes))!=num_nodes) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"only 2^k for now and MPI_COMM_WORLD!!! %D != %D\n",1<<i_log2_num_nodes,num_nodes);
92827bd09bSSatish Balay 
93827bd09bSSatish Balay   /* space for X info */
94a501084fSBarry Smith   xxt_handle->info = (xxt_info*)malloc(sizeof(xxt_info));
95827bd09bSSatish Balay 
96827bd09bSSatish Balay   /* set up matvec handles */
97827bd09bSSatish Balay   xxt_handle->mvi  = set_mvi(local2global, n, m, matvec, grid_data);
98827bd09bSSatish Balay 
99827bd09bSSatish Balay   /* matrix is assumed to be of full rank */
100827bd09bSSatish Balay   /* LATER we can reset to indicate rank def. */
101827bd09bSSatish Balay   xxt_handle->ns=0;
102827bd09bSSatish Balay 
103827bd09bSSatish Balay   /* determine separators and generate firing order - NB xxt info set here */
104827bd09bSSatish Balay   det_separators(xxt_handle);
105827bd09bSSatish Balay 
106827bd09bSSatish Balay   return(do_xxt_factor(xxt_handle));
107827bd09bSSatish Balay }
108827bd09bSSatish Balay 
1097b1ae94cSBarry Smith /**************************************xxt.c***********************************/
1108cda6cd7SBarry Smith PetscInt XXT_solve(xxt_ADT xxt_handle, PetscScalar *x, PetscScalar *b)
111827bd09bSSatish Balay {
112827bd09bSSatish Balay 
113a40341efSBarry Smith   comm_init();
114827bd09bSSatish Balay   check_handle(xxt_handle);
115827bd09bSSatish Balay 
116827bd09bSSatish Balay   /* need to copy b into x? */
117827bd09bSSatish Balay   if (b)
118827bd09bSSatish Balay     {rvec_copy(x,b,xxt_handle->mvi->n);}
119827bd09bSSatish Balay   do_xxt_solve(xxt_handle,x);
120827bd09bSSatish Balay 
121827bd09bSSatish Balay   return(0);
122827bd09bSSatish Balay }
123827bd09bSSatish Balay 
1247b1ae94cSBarry Smith /**************************************xxt.c***********************************/
12552f87cdaSBarry Smith PetscInt XXT_free(xxt_ADT xxt_handle)
126827bd09bSSatish Balay {
127827bd09bSSatish Balay 
128a40341efSBarry Smith   comm_init();
129827bd09bSSatish Balay   check_handle(xxt_handle);
130827bd09bSSatish Balay   n_xxt_handles--;
131827bd09bSSatish Balay 
132a501084fSBarry Smith   free(xxt_handle->info->nsep);
133a501084fSBarry Smith   free(xxt_handle->info->lnsep);
134a501084fSBarry Smith   free(xxt_handle->info->fo);
135a501084fSBarry Smith   free(xxt_handle->info->stages);
136a501084fSBarry Smith   free(xxt_handle->info->solve_uu);
137a501084fSBarry Smith   free(xxt_handle->info->solve_w);
138a501084fSBarry Smith   free(xxt_handle->info->x);
139a501084fSBarry Smith   free(xxt_handle->info->col_vals);
140a501084fSBarry Smith   free(xxt_handle->info->col_sz);
141a501084fSBarry Smith   free(xxt_handle->info->col_indices);
142a501084fSBarry Smith   free(xxt_handle->info);
143a501084fSBarry Smith   free(xxt_handle->mvi->local2global);
144827bd09bSSatish Balay    gs_free(xxt_handle->mvi->gs_handle);
145a501084fSBarry Smith   free(xxt_handle->mvi);
146a501084fSBarry Smith   free(xxt_handle);
147827bd09bSSatish Balay 
148827bd09bSSatish Balay   /* if the check fails we nuke */
149a501084fSBarry Smith   /* if NULL pointer passed to free we nuke */
150827bd09bSSatish Balay   /* if the calls to free fail that's not my problem */
151827bd09bSSatish Balay   return(0);
152827bd09bSSatish Balay }
153827bd09bSSatish Balay 
1547b1ae94cSBarry Smith /**************************************xxt.c***********************************/
15552f87cdaSBarry Smith PetscInt XXT_stats(xxt_ADT xxt_handle)
156827bd09bSSatish Balay {
15752f87cdaSBarry Smith   PetscInt       op[] = {NON_UNIFORM,GL_MIN,GL_MAX,GL_ADD,GL_MIN,GL_MAX,GL_ADD,GL_MIN,GL_MAX,GL_ADD};
15852f87cdaSBarry Smith   PetscInt       fop[] = {NON_UNIFORM,GL_MIN,GL_MAX,GL_ADD};
15952f87cdaSBarry Smith   PetscInt       vals[9],  work[9];
160a501084fSBarry Smith   PetscScalar    fvals[3], fwork[3];
16171a0148aSBarry Smith   PetscErrorCode ierr;
162827bd09bSSatish Balay 
163a40341efSBarry Smith   comm_init();
164827bd09bSSatish Balay   check_handle(xxt_handle);
165827bd09bSSatish Balay 
166827bd09bSSatish Balay   /* if factorization not done there are no stats */
167827bd09bSSatish Balay   if (!xxt_handle->info||!xxt_handle->mvi)
168827bd09bSSatish Balay     {
169*5a586d82SBarry Smith       if (!my_id) {ierr = PetscPrintf(PETSC_COMM_WORLD,"XXT_stats() :: no stats available!\n");CHKERRQ(ierr);}
170827bd09bSSatish Balay       return 1;
171827bd09bSSatish Balay     }
172827bd09bSSatish Balay 
173827bd09bSSatish Balay   vals[0]=vals[1]=vals[2]=xxt_handle->info->nnz;
174827bd09bSSatish Balay   vals[3]=vals[4]=vals[5]=xxt_handle->mvi->n;
175827bd09bSSatish Balay   vals[6]=vals[7]=vals[8]=xxt_handle->info->msg_buf_sz;
176827bd09bSSatish Balay   giop(vals,work,sizeof(op)/sizeof(op[0])-1,op);
177827bd09bSSatish Balay 
178827bd09bSSatish Balay   fvals[0]=fvals[1]=fvals[2]
179827bd09bSSatish Balay     =xxt_handle->info->tot_solve_time/xxt_handle->info->nsolves++;
180827bd09bSSatish Balay   grop(fvals,fwork,sizeof(fop)/sizeof(fop[0])-1,fop);
181827bd09bSSatish Balay 
182827bd09bSSatish Balay   if (!my_id)
183827bd09bSSatish Balay     {
184*5a586d82SBarry Smith       ierr = PetscPrintf(PETSC_COMM_WORLD,"%D :: min   xxt_nnz=%D\n",my_id,vals[0]);CHKERRQ(ierr);
185*5a586d82SBarry Smith       ierr = PetscPrintf(PETSC_COMM_WORLD,"%D :: max   xxt_nnz=%D\n",my_id,vals[1]);CHKERRQ(ierr);
186*5a586d82SBarry Smith       ierr = PetscPrintf(PETSC_COMM_WORLD,"%D :: avg   xxt_nnz=%g\n",my_id,1.0*vals[2]/num_nodes);CHKERRQ(ierr);
187*5a586d82SBarry Smith       ierr = PetscPrintf(PETSC_COMM_WORLD,"%D :: tot   xxt_nnz=%D\n",my_id,vals[2]);CHKERRQ(ierr);
188*5a586d82SBarry Smith       ierr = PetscPrintf(PETSC_COMM_WORLD,"%D :: xxt   C(2d)  =%g\n",my_id,vals[2]/(pow(1.0*vals[5],1.5)));CHKERRQ(ierr);
189*5a586d82SBarry Smith       ierr = PetscPrintf(PETSC_COMM_WORLD,"%D :: xxt   C(3d)  =%g\n",my_id,vals[2]/(pow(1.0*vals[5],1.6667)));CHKERRQ(ierr);
190*5a586d82SBarry Smith       ierr = PetscPrintf(PETSC_COMM_WORLD,"%D :: min   xxt_n  =%D\n",my_id,vals[3]);CHKERRQ(ierr);
191*5a586d82SBarry Smith       ierr = PetscPrintf(PETSC_COMM_WORLD,"%D :: max   xxt_n  =%D\n",my_id,vals[4]);CHKERRQ(ierr);
192*5a586d82SBarry Smith       ierr = PetscPrintf(PETSC_COMM_WORLD,"%D :: avg   xxt_n  =%g\n",my_id,1.0*vals[5]/num_nodes);CHKERRQ(ierr);
193*5a586d82SBarry Smith       ierr = PetscPrintf(PETSC_COMM_WORLD,"%D :: tot   xxt_n  =%D\n",my_id,vals[5]);CHKERRQ(ierr);
194*5a586d82SBarry Smith       ierr = PetscPrintf(PETSC_COMM_WORLD,"%D :: min   xxt_buf=%D\n",my_id,vals[6]);CHKERRQ(ierr);
195*5a586d82SBarry Smith       ierr = PetscPrintf(PETSC_COMM_WORLD,"%D :: max   xxt_buf=%D\n",my_id,vals[7]);CHKERRQ(ierr);
196*5a586d82SBarry Smith       ierr = PetscPrintf(PETSC_COMM_WORLD,"%D :: avg   xxt_buf=%g\n",my_id,1.0*vals[8]/num_nodes);CHKERRQ(ierr);
197*5a586d82SBarry Smith       ierr = PetscPrintf(PETSC_COMM_WORLD,"%D :: min   xxt_slv=%g\n",my_id,fvals[0]);CHKERRQ(ierr);
198*5a586d82SBarry Smith       ierr = PetscPrintf(PETSC_COMM_WORLD,"%D :: max   xxt_slv=%g\n",my_id,fvals[1]);CHKERRQ(ierr);
199*5a586d82SBarry Smith       ierr = PetscPrintf(PETSC_COMM_WORLD,"%D :: avg   xxt_slv=%g\n",my_id,fvals[2]/num_nodes);CHKERRQ(ierr);
200827bd09bSSatish Balay     }
201827bd09bSSatish Balay 
202827bd09bSSatish Balay   return(0);
203827bd09bSSatish Balay }
204827bd09bSSatish Balay 
205827bd09bSSatish Balay /*************************************xxt.c************************************
206827bd09bSSatish Balay 
207827bd09bSSatish Balay Description: get A_local, local portion of global coarse matrix which
208827bd09bSSatish Balay is a row dist. nxm matrix w/ n<m.
209827bd09bSSatish Balay    o my_ml holds address of ML struct associated w/A_local and coarse grid
210827bd09bSSatish Balay    o local2global holds global number of column i (i=0,...,m-1)
211827bd09bSSatish Balay    o local2global holds global number of row    i (i=0,...,n-1)
212827bd09bSSatish Balay    o mylocmatvec performs A_local . vec_local (note that gs is performed using
213827bd09bSSatish Balay    gs_init/gop).
214827bd09bSSatish Balay 
215827bd09bSSatish Balay mylocmatvec = my_ml->Amat[grid_tag].matvec->external;
216827bd09bSSatish Balay mylocmatvec (void :: void *data, double *in, double *out)
217827bd09bSSatish Balay **************************************xxt.c***********************************/
21852f87cdaSBarry Smith static PetscInt do_xxt_factor(xxt_ADT xxt_handle)
219827bd09bSSatish Balay {
2207b1ae94cSBarry Smith   return xxt_generate(xxt_handle);
221827bd09bSSatish Balay }
222827bd09bSSatish Balay 
2237b1ae94cSBarry Smith /**************************************xxt.c***********************************/
22452f87cdaSBarry Smith static PetscInt xxt_generate(xxt_ADT xxt_handle)
225827bd09bSSatish Balay {
22652f87cdaSBarry Smith   PetscInt i,j,k,idex;
22752f87cdaSBarry Smith   PetscInt dim, col;
228a501084fSBarry Smith   PetscScalar *u, *uu, *v, *z, *w, alpha, alpha_w;
22952f87cdaSBarry Smith   PetscInt *segs;
23052f87cdaSBarry Smith   PetscInt op[] = {GL_ADD,0};
23152f87cdaSBarry Smith   PetscInt off, len;
232a501084fSBarry Smith   PetscScalar *x_ptr;
23352f87cdaSBarry Smith   PetscInt *iptr, flag;
23452f87cdaSBarry Smith   PetscInt start=0, end, work;
23552f87cdaSBarry Smith   PetscInt op2[] = {GL_MIN,0};
236827bd09bSSatish Balay   gs_ADT gs_handle;
23752f87cdaSBarry Smith   PetscInt *nsep, *lnsep, *fo;
23852f87cdaSBarry Smith   PetscInt a_n=xxt_handle->mvi->n;
23952f87cdaSBarry Smith   PetscInt a_m=xxt_handle->mvi->m;
24052f87cdaSBarry Smith   PetscInt *a_local2global=xxt_handle->mvi->local2global;
24152f87cdaSBarry Smith   PetscInt level;
24252f87cdaSBarry Smith   PetscInt xxt_nnz=0, xxt_max_nnz=0;
24352f87cdaSBarry Smith   PetscInt n, m;
24452f87cdaSBarry Smith   PetscInt *col_sz, *col_indices, *stages;
245a501084fSBarry Smith   PetscScalar **col_vals, *x;
24652f87cdaSBarry Smith   PetscInt n_global;
24752f87cdaSBarry Smith   PetscInt xxt_zero_nnz=0;
24852f87cdaSBarry Smith   PetscInt xxt_zero_nnz_0=0;
24971a0148aSBarry Smith   PetscBLASInt i1 = 1,dlen;
250a501084fSBarry Smith   PetscScalar dm1 = -1.0;
251d1528f56SBarry Smith   PetscErrorCode ierr;
252827bd09bSSatish Balay 
253827bd09bSSatish Balay   n=xxt_handle->mvi->n;
254827bd09bSSatish Balay   nsep=xxt_handle->info->nsep;
255827bd09bSSatish Balay   lnsep=xxt_handle->info->lnsep;
256827bd09bSSatish Balay   fo=xxt_handle->info->fo;
257827bd09bSSatish Balay   end=lnsep[0];
258827bd09bSSatish Balay   level=xxt_handle->level;
259827bd09bSSatish Balay   gs_handle=xxt_handle->mvi->gs_handle;
260827bd09bSSatish Balay 
261827bd09bSSatish Balay   /* is there a null space? */
262827bd09bSSatish Balay   /* LATER add in ability to detect null space by checking alpha */
263827bd09bSSatish Balay   for (i=0, j=0; i<=level; i++)
264827bd09bSSatish Balay     {j+=nsep[i];}
265827bd09bSSatish Balay 
266827bd09bSSatish Balay   m = j-xxt_handle->ns;
267827bd09bSSatish Balay   if (m!=j)
26871a0148aSBarry Smith     {ierr = PetscPrintf(PETSC_COMM_WORLD,"xxt_generate() :: null space exists %D %D %D\n",m,j,xxt_handle->ns);}
269827bd09bSSatish Balay 
270827bd09bSSatish Balay   /* get and initialize storage for x local         */
271827bd09bSSatish Balay   /* note that x local is nxm and stored by columns */
27252f87cdaSBarry Smith   col_sz = (PetscInt*) malloc(m*sizeof(PetscInt));
27352f87cdaSBarry Smith   col_indices = (PetscInt*) malloc((2*m+1)*sizeof(PetscInt));
274a501084fSBarry Smith   col_vals = (PetscScalar **) malloc(m*sizeof(PetscScalar *));
275827bd09bSSatish Balay   for (i=j=0; i<m; i++, j+=2)
276827bd09bSSatish Balay     {
277827bd09bSSatish Balay       col_indices[j]=col_indices[j+1]=col_sz[i]=-1;
278827bd09bSSatish Balay       col_vals[i] = NULL;
279827bd09bSSatish Balay     }
280827bd09bSSatish Balay   col_indices[j]=-1;
281827bd09bSSatish Balay 
282827bd09bSSatish Balay   /* size of separators for each sub-hc working from bottom of tree to top */
283827bd09bSSatish Balay   /* this looks like nsep[]=segments */
28452f87cdaSBarry Smith   stages = (PetscInt*) malloc((level+1)*sizeof(PetscInt));
28552f87cdaSBarry Smith   segs   = (PetscInt*) malloc((level+1)*sizeof(PetscInt));
286827bd09bSSatish Balay   ivec_zero(stages,level+1);
287827bd09bSSatish Balay   ivec_copy(segs,nsep,level+1);
288827bd09bSSatish Balay   for (i=0; i<level; i++)
289827bd09bSSatish Balay     {segs[i+1] += segs[i];}
290827bd09bSSatish Balay   stages[0] = segs[0];
291827bd09bSSatish Balay 
292827bd09bSSatish Balay   /* temporary vectors  */
293a501084fSBarry Smith   u  = (PetscScalar *) malloc(n*sizeof(PetscScalar));
294a501084fSBarry Smith   z  = (PetscScalar *) malloc(n*sizeof(PetscScalar));
295a501084fSBarry Smith   v  = (PetscScalar *) malloc(a_m*sizeof(PetscScalar));
296a501084fSBarry Smith   uu = (PetscScalar *) malloc(m*sizeof(PetscScalar));
297a501084fSBarry Smith   w  = (PetscScalar *) malloc(m*sizeof(PetscScalar));
298827bd09bSSatish Balay 
299827bd09bSSatish Balay   /* extra nnz due to replication of vertices across separators */
300827bd09bSSatish Balay   for (i=1, j=0; i<=level; i++)
301827bd09bSSatish Balay     {j+=nsep[i];}
302827bd09bSSatish Balay 
303827bd09bSSatish Balay   /* storage for sparse x values */
304827bd09bSSatish Balay   n_global = xxt_handle->info->n_global;
30552f87cdaSBarry Smith   xxt_max_nnz = (PetscInt)(2.5*pow(1.0*n_global,1.6667) + j*n/2)/num_nodes;
306a501084fSBarry Smith   x = (PetscScalar *) malloc(xxt_max_nnz*sizeof(PetscScalar));
307827bd09bSSatish Balay   xxt_nnz = 0;
308827bd09bSSatish Balay 
309827bd09bSSatish Balay   /* LATER - can embed next sep to fire in gs */
310827bd09bSSatish Balay   /* time to make the donuts - generate X factor */
311827bd09bSSatish Balay   for (dim=i=j=0;i<m;i++)
312827bd09bSSatish Balay     {
313827bd09bSSatish Balay       /* time to move to the next level? */
314d4af0d30SBarry Smith       while (i==segs[dim]) {
315e32f2f54SBarry Smith         if (dim==level) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"dim about to exceed level\n");
316827bd09bSSatish Balay         stages[dim++]=i;
317827bd09bSSatish Balay         end+=lnsep[dim];
318827bd09bSSatish Balay       }
319827bd09bSSatish Balay       stages[dim]=i;
320827bd09bSSatish Balay 
321827bd09bSSatish Balay       /* which column are we firing? */
322827bd09bSSatish Balay       /* i.e. set v_l */
323827bd09bSSatish Balay       /* use new seps and do global min across hc to determine which one to fire */
324827bd09bSSatish Balay       (start<end) ? (col=fo[start]) : (col=INT_MAX);
325827bd09bSSatish Balay       giop_hc(&col,&work,1,op2,dim);
326827bd09bSSatish Balay 
327827bd09bSSatish Balay       /* shouldn't need this */
328827bd09bSSatish Balay       if (col==INT_MAX)
329827bd09bSSatish Balay 	{
330f1ed62a8SBarry Smith 	  ierr = PetscInfo(0,"hey ... col==INT_MAX??\n");CHKERRQ(ierr);
331827bd09bSSatish Balay 	  continue;
332827bd09bSSatish Balay 	}
333827bd09bSSatish Balay 
334827bd09bSSatish Balay       /* do I own it? I should */
335827bd09bSSatish Balay       rvec_zero(v ,a_m);
336827bd09bSSatish Balay       if (col==fo[start])
337827bd09bSSatish Balay 	{
338827bd09bSSatish Balay 	  start++;
339827bd09bSSatish Balay 	  idex=ivec_linear_search(col, a_local2global, a_n);
340e7e72b3dSBarry Smith 	  if (idex!=-1){
341e7e72b3dSBarry Smith             v[idex] = 1.0; j++;
342e7e72b3dSBarry Smith           } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"NOT FOUND!\n");
343827bd09bSSatish Balay 	}
344827bd09bSSatish Balay       else
345827bd09bSSatish Balay 	{
346827bd09bSSatish Balay 	  idex=ivec_linear_search(col, a_local2global, a_m);
347827bd09bSSatish Balay 	  if (idex!=-1)
348827bd09bSSatish Balay 	    {v[idex] = 1.0;}
349827bd09bSSatish Balay 	}
350827bd09bSSatish Balay 
351827bd09bSSatish Balay       /* perform u = A.v_l */
352827bd09bSSatish Balay       rvec_zero(u,n);
353827bd09bSSatish Balay       do_matvec(xxt_handle->mvi,v,u);
354827bd09bSSatish Balay 
355827bd09bSSatish Balay       /* uu =  X^T.u_l (local portion) */
356827bd09bSSatish Balay       /* technically only need to zero out first i entries */
357827bd09bSSatish Balay       /* later turn this into an XXT_solve call ? */
358827bd09bSSatish Balay       rvec_zero(uu,m);
359827bd09bSSatish Balay       x_ptr=x;
360827bd09bSSatish Balay       iptr = col_indices;
361827bd09bSSatish Balay       for (k=0; k<i; k++)
362827bd09bSSatish Balay 	{
363827bd09bSSatish Balay 	  off = *iptr++;
364827bd09bSSatish Balay 	  len = *iptr++;
3650805154bSBarry Smith           dlen = PetscBLASIntCast(len);
36671a0148aSBarry Smith 	  uu[k] = BLASdot_(&dlen,u+off,&i1,x_ptr,&i1);
367827bd09bSSatish Balay 	  x_ptr+=len;
368827bd09bSSatish Balay 	}
369827bd09bSSatish Balay 
370827bd09bSSatish Balay 
371827bd09bSSatish Balay       /* uu = X^T.u_l (comm portion) */
372827bd09bSSatish Balay       ssgl_radd  (uu, w, dim, stages);
373827bd09bSSatish Balay 
374827bd09bSSatish Balay       /* z = X.uu */
375827bd09bSSatish Balay       rvec_zero(z,n);
376827bd09bSSatish Balay       x_ptr=x;
377827bd09bSSatish Balay       iptr = col_indices;
378827bd09bSSatish Balay       for (k=0; k<i; k++)
379827bd09bSSatish Balay 	{
380827bd09bSSatish Balay 	  off = *iptr++;
381827bd09bSSatish Balay 	  len = *iptr++;
3820805154bSBarry Smith           dlen = PetscBLASIntCast(len);
38371a0148aSBarry Smith 	  BLASaxpy_(&dlen,&uu[k],x_ptr,&i1,z+off,&i1);
384827bd09bSSatish Balay 	  x_ptr+=len;
385827bd09bSSatish Balay 	}
386827bd09bSSatish Balay 
387827bd09bSSatish Balay       /* compute v_l = v_l - z */
388827bd09bSSatish Balay       rvec_zero(v+a_n,a_m-a_n);
3890805154bSBarry Smith       dlen = PetscBLASIntCast(n);
39071a0148aSBarry Smith       BLASaxpy_(&dlen,&dm1,z,&i1,v,&i1);
391827bd09bSSatish Balay 
392827bd09bSSatish Balay       /* compute u_l = A.v_l */
393827bd09bSSatish Balay       if (a_n!=a_m)
394827bd09bSSatish Balay 	{gs_gop_hc(gs_handle,v,"+\0",dim);}
395827bd09bSSatish Balay       rvec_zero(u,n);
396827bd09bSSatish Balay       do_matvec(xxt_handle->mvi,v,u);
397827bd09bSSatish Balay 
398827bd09bSSatish Balay       /* compute sqrt(alpha) = sqrt(v_l^T.u_l) - local portion */
3990805154bSBarry Smith       dlen = PetscBLASIntCast(n);
40071a0148aSBarry Smith       alpha = BLASdot_(&dlen,u,&i1,v,&i1);
401827bd09bSSatish Balay       /* compute sqrt(alpha) = sqrt(v_l^T.u_l) - comm portion */
402827bd09bSSatish Balay       grop_hc(&alpha, &alpha_w, 1, op, dim);
403827bd09bSSatish Balay 
404a501084fSBarry Smith       alpha = (PetscScalar) sqrt((double)alpha);
405827bd09bSSatish Balay 
406827bd09bSSatish Balay       /* check for small alpha                             */
407827bd09bSSatish Balay       /* LATER use this to detect and determine null space */
408c1235816SBarry Smith       if (fabs(alpha)<1.0e-14) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_PLIB,"bad alpha! %g\n",alpha);
409827bd09bSSatish Balay 
410827bd09bSSatish Balay       /* compute v_l = v_l/sqrt(alpha) */
411827bd09bSSatish Balay       rvec_scale(v,1.0/alpha,n);
412827bd09bSSatish Balay 
413827bd09bSSatish Balay       /* add newly generated column, v_l, to X */
414827bd09bSSatish Balay       flag = 1;
415827bd09bSSatish Balay       off=len=0;
416827bd09bSSatish Balay       for (k=0; k<n; k++)
417827bd09bSSatish Balay 	{
418827bd09bSSatish Balay 	  if (v[k]!=0.0)
419827bd09bSSatish Balay 	    {
420827bd09bSSatish Balay 	      len=k;
421827bd09bSSatish Balay 	      if (flag)
422827bd09bSSatish Balay 		{off=k; flag=0;}
423827bd09bSSatish Balay 	    }
424827bd09bSSatish Balay 	}
425827bd09bSSatish Balay 
426827bd09bSSatish Balay       len -= (off-1);
427827bd09bSSatish Balay 
428827bd09bSSatish Balay       if (len>0)
429827bd09bSSatish Balay 	{
430827bd09bSSatish Balay 	  if ((xxt_nnz+len)>xxt_max_nnz)
431827bd09bSSatish Balay 	    {
432f1ed62a8SBarry Smith 	      ierr = PetscInfo(0,"increasing space for X by 2x!\n");CHKERRQ(ierr);
433827bd09bSSatish Balay 	      xxt_max_nnz *= 2;
434a501084fSBarry Smith 	      x_ptr = (PetscScalar *) malloc(xxt_max_nnz*sizeof(PetscScalar));
435827bd09bSSatish Balay 	      rvec_copy(x_ptr,x,xxt_nnz);
436a501084fSBarry Smith 	      free(x);
437827bd09bSSatish Balay 	      x = x_ptr;
438827bd09bSSatish Balay 	      x_ptr+=xxt_nnz;
439827bd09bSSatish Balay 	    }
440827bd09bSSatish Balay 	  xxt_nnz += len;
441827bd09bSSatish Balay 	  rvec_copy(x_ptr,v+off,len);
442827bd09bSSatish Balay 
443827bd09bSSatish Balay           /* keep track of number of zeros */
444827bd09bSSatish Balay 	  if (dim)
445827bd09bSSatish Balay 	    {
446827bd09bSSatish Balay 	      for (k=0; k<len; k++)
447827bd09bSSatish Balay 		{
448827bd09bSSatish Balay 		  if (x_ptr[k]==0.0)
449827bd09bSSatish Balay 		    {xxt_zero_nnz++;}
450827bd09bSSatish Balay 		}
451827bd09bSSatish Balay 	    }
452827bd09bSSatish Balay 	  else
453827bd09bSSatish Balay 	    {
454827bd09bSSatish Balay 	      for (k=0; k<len; k++)
455827bd09bSSatish Balay 		{
456827bd09bSSatish Balay 		  if (x_ptr[k]==0.0)
457827bd09bSSatish Balay 		    {xxt_zero_nnz_0++;}
458827bd09bSSatish Balay 		}
459827bd09bSSatish Balay 	    }
460827bd09bSSatish Balay 	  col_indices[2*i] = off;
461827bd09bSSatish Balay 	  col_sz[i] = col_indices[2*i+1] = len;
462827bd09bSSatish Balay 	  col_vals[i] = x_ptr;
463827bd09bSSatish Balay 	}
464827bd09bSSatish Balay       else
465827bd09bSSatish Balay 	{
466827bd09bSSatish Balay 	  col_indices[2*i] = 0;
467827bd09bSSatish Balay 	  col_sz[i] = col_indices[2*i+1] = 0;
468827bd09bSSatish Balay 	  col_vals[i] = x_ptr;
469827bd09bSSatish Balay 	}
470827bd09bSSatish Balay     }
471827bd09bSSatish Balay 
472827bd09bSSatish Balay   /* close off stages for execution phase */
473827bd09bSSatish Balay   while (dim!=level)
474827bd09bSSatish Balay     {
475827bd09bSSatish Balay       stages[dim++]=i;
47671a0148aSBarry Smith       ierr = PetscInfo2(0,"disconnected!!! dim(%D)!=level(%D)\n",dim,level);CHKERRQ(ierr);
477827bd09bSSatish Balay     }
478827bd09bSSatish Balay   stages[dim]=i;
479827bd09bSSatish Balay 
480827bd09bSSatish Balay   xxt_handle->info->n=xxt_handle->mvi->n;
481827bd09bSSatish Balay   xxt_handle->info->m=m;
482827bd09bSSatish Balay   xxt_handle->info->nnz=xxt_nnz;
483827bd09bSSatish Balay   xxt_handle->info->max_nnz=xxt_max_nnz;
484827bd09bSSatish Balay   xxt_handle->info->msg_buf_sz=stages[level]-stages[0];
485a501084fSBarry Smith   xxt_handle->info->solve_uu = (PetscScalar *) malloc(m*sizeof(PetscScalar));
486a501084fSBarry Smith   xxt_handle->info->solve_w  = (PetscScalar *) malloc(m*sizeof(PetscScalar));
487827bd09bSSatish Balay   xxt_handle->info->x=x;
488827bd09bSSatish Balay   xxt_handle->info->col_vals=col_vals;
489827bd09bSSatish Balay   xxt_handle->info->col_sz=col_sz;
490827bd09bSSatish Balay   xxt_handle->info->col_indices=col_indices;
491827bd09bSSatish Balay   xxt_handle->info->stages=stages;
492827bd09bSSatish Balay   xxt_handle->info->nsolves=0;
493827bd09bSSatish Balay   xxt_handle->info->tot_solve_time=0.0;
494827bd09bSSatish Balay 
495a501084fSBarry Smith   free(segs);
496a501084fSBarry Smith   free(u);
497a501084fSBarry Smith   free(v);
498a501084fSBarry Smith   free(uu);
499a501084fSBarry Smith   free(z);
500a501084fSBarry Smith   free(w);
501827bd09bSSatish Balay 
502827bd09bSSatish Balay   return(0);
503827bd09bSSatish Balay }
504827bd09bSSatish Balay 
5057b1ae94cSBarry Smith /**************************************xxt.c***********************************/
5060924e98cSBarry Smith static PetscErrorCode do_xxt_solve(xxt_ADT xxt_handle,  PetscScalar *uc)
507827bd09bSSatish Balay {
50852f87cdaSBarry Smith    PetscInt off, len, *iptr;
50952f87cdaSBarry Smith   PetscInt level       =xxt_handle->level;
51052f87cdaSBarry Smith   PetscInt n           =xxt_handle->info->n;
51152f87cdaSBarry Smith   PetscInt m           =xxt_handle->info->m;
51252f87cdaSBarry Smith   PetscInt *stages     =xxt_handle->info->stages;
51352f87cdaSBarry Smith   PetscInt *col_indices=xxt_handle->info->col_indices;
514a501084fSBarry Smith   PetscScalar *x_ptr, *uu_ptr;
515a501084fSBarry Smith   PetscScalar *solve_uu=xxt_handle->info->solve_uu;
516a501084fSBarry Smith   PetscScalar *solve_w =xxt_handle->info->solve_w;
517a501084fSBarry Smith   PetscScalar *x       =xxt_handle->info->x;
51871a0148aSBarry Smith   PetscBLASInt i1 = 1,dlen;
519827bd09bSSatish Balay 
5200924e98cSBarry Smith   PetscFunctionBegin;
521827bd09bSSatish Balay   uu_ptr=solve_uu;
522827bd09bSSatish Balay   rvec_zero(uu_ptr,m);
523827bd09bSSatish Balay 
524827bd09bSSatish Balay   /* x  = X.Y^T.b */
525827bd09bSSatish Balay   /* uu = Y^T.b */
526827bd09bSSatish Balay   for (x_ptr=x,iptr=col_indices; *iptr!=-1; x_ptr+=len)
527827bd09bSSatish Balay     {
528827bd09bSSatish Balay       off=*iptr++; len=*iptr++;
5290805154bSBarry Smith       dlen = PetscBLASIntCast(len);
53071a0148aSBarry Smith       *uu_ptr++ = BLASdot_(&dlen,uc+off,&i1,x_ptr,&i1);
531827bd09bSSatish Balay     }
532827bd09bSSatish Balay 
533827bd09bSSatish Balay   /* comunication of beta */
534827bd09bSSatish Balay   uu_ptr=solve_uu;
535827bd09bSSatish Balay   if (level) {ssgl_radd(uu_ptr, solve_w, level, stages);}
536827bd09bSSatish Balay 
537827bd09bSSatish Balay   rvec_zero(uc,n);
538827bd09bSSatish Balay 
539827bd09bSSatish Balay   /* x = X.uu */
540827bd09bSSatish Balay   for (x_ptr=x,iptr=col_indices; *iptr!=-1; x_ptr+=len)
541827bd09bSSatish Balay     {
542827bd09bSSatish Balay       off=*iptr++; len=*iptr++;
5430805154bSBarry Smith       dlen = PetscBLASIntCast(len);
54471a0148aSBarry Smith       BLASaxpy_(&dlen,uu_ptr++,x_ptr,&i1,uc+off,&i1);
545827bd09bSSatish Balay     }
5460924e98cSBarry Smith   PetscFunctionReturn(0);
547827bd09bSSatish Balay }
548827bd09bSSatish Balay 
5497b1ae94cSBarry Smith /**************************************xxt.c***********************************/
5500924e98cSBarry Smith static PetscErrorCode check_handle(xxt_ADT xxt_handle)
551827bd09bSSatish Balay {
55252f87cdaSBarry Smith   PetscInt vals[2], work[2], op[] = {NON_UNIFORM,GL_MIN,GL_MAX};
553827bd09bSSatish Balay 
5540924e98cSBarry Smith   PetscFunctionBegin;
555e32f2f54SBarry Smith   if (!xxt_handle) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_PLIB,"check_handle() :: bad handle :: NULL %D\n",xxt_handle);
556827bd09bSSatish Balay 
557827bd09bSSatish Balay   vals[0]=vals[1]=xxt_handle->id;
558827bd09bSSatish Balay   giop(vals,work,sizeof(op)/sizeof(op[0])-1,op);
559e32f2f54SBarry Smith   if ((vals[0]!=vals[1])||(xxt_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], xxt_handle->id);
5600924e98cSBarry Smith   PetscFunctionReturn(0);
561827bd09bSSatish Balay }
562827bd09bSSatish Balay 
5637b1ae94cSBarry Smith /**************************************xxt.c***********************************/
5640924e98cSBarry Smith static  PetscErrorCode det_separators(xxt_ADT xxt_handle)
565827bd09bSSatish Balay {
56652f87cdaSBarry Smith   PetscInt i, ct, id;
56752f87cdaSBarry Smith   PetscInt mask, edge, *iptr;
56852f87cdaSBarry Smith   PetscInt *dir, *used;
56952f87cdaSBarry Smith   PetscInt sum[4], w[4];
570a501084fSBarry Smith   PetscScalar rsum[4], rw[4];
57152f87cdaSBarry Smith   PetscInt op[] = {GL_ADD,0};
572a501084fSBarry Smith   PetscScalar *lhs, *rhs;
57352f87cdaSBarry Smith   PetscInt *nsep, *lnsep, *fo, nfo=0;
574827bd09bSSatish Balay   gs_ADT gs_handle=xxt_handle->mvi->gs_handle;
57552f87cdaSBarry Smith   PetscInt *local2global=xxt_handle->mvi->local2global;
57652f87cdaSBarry Smith   PetscInt  n=xxt_handle->mvi->n;
57752f87cdaSBarry Smith   PetscInt  m=xxt_handle->mvi->m;
57852f87cdaSBarry Smith   PetscInt level=xxt_handle->level;
57952f87cdaSBarry Smith   PetscInt shared=FALSE;
580827bd09bSSatish Balay 
5810924e98cSBarry Smith   PetscFunctionBegin;
58252f87cdaSBarry Smith   dir  = (PetscInt*)malloc(sizeof(PetscInt)*(level+1));
58352f87cdaSBarry Smith   nsep = (PetscInt*)malloc(sizeof(PetscInt)*(level+1));
58452f87cdaSBarry Smith   lnsep= (PetscInt*)malloc(sizeof(PetscInt)*(level+1));
58552f87cdaSBarry Smith   fo   = (PetscInt*)malloc(sizeof(PetscInt)*(n+1));
58652f87cdaSBarry Smith   used = (PetscInt*)malloc(sizeof(PetscInt)*n);
587827bd09bSSatish Balay 
588827bd09bSSatish Balay   ivec_zero(dir  ,level+1);
589827bd09bSSatish Balay   ivec_zero(nsep ,level+1);
590827bd09bSSatish Balay   ivec_zero(lnsep,level+1);
591827bd09bSSatish Balay   ivec_set (fo   ,-1,n+1);
592827bd09bSSatish Balay   ivec_zero(used,n);
593827bd09bSSatish Balay 
5948cda6cd7SBarry Smith   lhs  = (PetscScalar*)malloc(sizeof(PetscScalar)*m);
5958cda6cd7SBarry Smith   rhs  = (PetscScalar*)malloc(sizeof(PetscScalar)*m);
596827bd09bSSatish Balay 
597827bd09bSSatish Balay   /* determine the # of unique dof */
598827bd09bSSatish Balay   rvec_zero(lhs,m);
599827bd09bSSatish Balay   rvec_set(lhs,1.0,n);
600827bd09bSSatish Balay   gs_gop_hc(gs_handle,lhs,"+\0",level);
601827bd09bSSatish Balay   rvec_zero(rsum,2);
602827bd09bSSatish Balay   for (ct=i=0;i<n;i++)
603827bd09bSSatish Balay     {
604827bd09bSSatish Balay       if (lhs[i]!=0.0)
605827bd09bSSatish Balay 	{rsum[0]+=1.0/lhs[i]; rsum[1]+=lhs[i];}
606827bd09bSSatish Balay     }
607827bd09bSSatish Balay   grop_hc(rsum,rw,2,op,level);
608827bd09bSSatish Balay   rsum[0]+=0.1;
609827bd09bSSatish Balay   rsum[1]+=0.1;
610827bd09bSSatish Balay 
611827bd09bSSatish Balay   if (fabs(rsum[0]-rsum[1])>EPS)
612827bd09bSSatish Balay     {shared=TRUE;}
613827bd09bSSatish Balay 
61452f87cdaSBarry Smith   xxt_handle->info->n_global=xxt_handle->info->m_global=(PetscInt) rsum[0];
61552f87cdaSBarry Smith   xxt_handle->mvi->n_global =xxt_handle->mvi->m_global =(PetscInt) rsum[0];
616827bd09bSSatish Balay 
617827bd09bSSatish Balay   /* determine separator sets top down */
618827bd09bSSatish Balay   if (shared)
619827bd09bSSatish Balay     {
620827bd09bSSatish Balay       for (iptr=fo+n,id=my_id,mask=num_nodes>>1,edge=level;edge>0;edge--,mask>>=1)
621827bd09bSSatish Balay 	{
622827bd09bSSatish Balay 	  /* set rsh of hc, fire, and collect lhs responses */
623827bd09bSSatish Balay 	  (id<mask) ? rvec_zero(lhs,m) : rvec_set(lhs,1.0,m);
624827bd09bSSatish Balay 	  gs_gop_hc(gs_handle,lhs,"+\0",edge);
625827bd09bSSatish Balay 
626827bd09bSSatish Balay 	  /* set lsh of hc, fire, and collect rhs responses */
627827bd09bSSatish Balay 	  (id<mask) ? rvec_set(rhs,1.0,m) : rvec_zero(rhs,m);
628827bd09bSSatish Balay 	  gs_gop_hc(gs_handle,rhs,"+\0",edge);
629827bd09bSSatish Balay 
630827bd09bSSatish Balay 	  for (i=0;i<n;i++)
631827bd09bSSatish Balay 	    {
632827bd09bSSatish Balay 	      if (id< mask)
633827bd09bSSatish Balay 		{
634827bd09bSSatish Balay 		  if (lhs[i]!=0.0)
635827bd09bSSatish Balay 		    {lhs[i]=1.0;}
636827bd09bSSatish Balay 		}
637827bd09bSSatish Balay 	      if (id>=mask)
638827bd09bSSatish Balay 		{
639827bd09bSSatish Balay 		  if (rhs[i]!=0.0)
640827bd09bSSatish Balay 		    {rhs[i]=1.0;}
641827bd09bSSatish Balay 		}
642827bd09bSSatish Balay 	    }
643827bd09bSSatish Balay 
644827bd09bSSatish Balay 	  if (id< mask)
645827bd09bSSatish Balay 	    {gs_gop_hc(gs_handle,lhs,"+\0",edge-1);}
646827bd09bSSatish Balay 	  else
647827bd09bSSatish Balay 	    {gs_gop_hc(gs_handle,rhs,"+\0",edge-1);}
648827bd09bSSatish Balay 
649827bd09bSSatish Balay 	  /* count number of dofs I own that have signal and not in sep set */
650827bd09bSSatish Balay 	  rvec_zero(rsum,4);
651827bd09bSSatish Balay 	  for (ivec_zero(sum,4),ct=i=0;i<n;i++)
652827bd09bSSatish Balay 	    {
653827bd09bSSatish Balay 	      if (!used[i])
654827bd09bSSatish Balay 		{
655827bd09bSSatish Balay 		  /* number of unmarked dofs on node */
656827bd09bSSatish Balay 		  ct++;
657827bd09bSSatish Balay 		  /* number of dofs to be marked on lhs hc */
658827bd09bSSatish Balay 		  if (id< mask)
659827bd09bSSatish Balay 		    {
660827bd09bSSatish Balay 		      if (lhs[i]!=0.0)
661827bd09bSSatish Balay 			{sum[0]++; rsum[0]+=1.0/lhs[i];}
662827bd09bSSatish Balay 		    }
663827bd09bSSatish Balay 		  /* number of dofs to be marked on rhs hc */
664827bd09bSSatish Balay 		  if (id>=mask)
665827bd09bSSatish Balay 		    {
666827bd09bSSatish Balay 		      if (rhs[i]!=0.0)
667827bd09bSSatish Balay 			{sum[1]++; rsum[1]+=1.0/rhs[i];}
668827bd09bSSatish Balay 		    }
669827bd09bSSatish Balay 		}
670827bd09bSSatish Balay 	    }
671827bd09bSSatish Balay 
672827bd09bSSatish Balay 	  /* go for load balance - choose half with most unmarked dofs, bias LHS */
673827bd09bSSatish Balay 	  (id<mask) ? (sum[2]=ct) : (sum[3]=ct);
674827bd09bSSatish Balay 	  (id<mask) ? (rsum[2]=ct) : (rsum[3]=ct);
675827bd09bSSatish Balay 	  giop_hc(sum,w,4,op,edge);
676827bd09bSSatish Balay 	  grop_hc(rsum,rw,4,op,edge);
677827bd09bSSatish Balay 	  rsum[0]+=0.1; rsum[1]+=0.1; rsum[2]+=0.1; rsum[3]+=0.1;
678827bd09bSSatish Balay 
679827bd09bSSatish Balay 	  if (id<mask)
680827bd09bSSatish Balay 	    {
681827bd09bSSatish Balay 	      /* mark dofs I own that have signal and not in sep set */
682827bd09bSSatish Balay 	      for (ct=i=0;i<n;i++)
683827bd09bSSatish Balay 		{
684827bd09bSSatish Balay 		  if ((!used[i])&&(lhs[i]!=0.0))
685827bd09bSSatish Balay 		    {
686827bd09bSSatish Balay 		      ct++; nfo++;
687827bd09bSSatish Balay 
688c1235816SBarry Smith 		      if (nfo>n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"nfo about to exceed n\n");
689827bd09bSSatish Balay 
690827bd09bSSatish Balay 		      *--iptr = local2global[i];
691827bd09bSSatish Balay 		      used[i]=edge;
692827bd09bSSatish Balay 		    }
693827bd09bSSatish Balay 		}
694827bd09bSSatish Balay 	      if (ct>1) {ivec_sort(iptr,ct);}
695827bd09bSSatish Balay 
696827bd09bSSatish Balay 	      lnsep[edge]=ct;
69752f87cdaSBarry Smith 	      nsep[edge]=(PetscInt) rsum[0];
698827bd09bSSatish Balay 	      dir [edge]=LEFT;
699827bd09bSSatish Balay 	    }
700827bd09bSSatish Balay 
701827bd09bSSatish Balay 	  if (id>=mask)
702827bd09bSSatish Balay 	    {
703827bd09bSSatish Balay 	      /* mark dofs I own that have signal and not in sep set */
704827bd09bSSatish Balay 	      for (ct=i=0;i<n;i++)
705827bd09bSSatish Balay 		{
706827bd09bSSatish Balay 		  if ((!used[i])&&(rhs[i]!=0.0))
707827bd09bSSatish Balay 		    {
708827bd09bSSatish Balay 		      ct++; nfo++;
709827bd09bSSatish Balay 
710c1235816SBarry Smith 		      if (nfo>n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"nfo about to exceed n\n");
711827bd09bSSatish Balay 
712827bd09bSSatish Balay 		      *--iptr = local2global[i];
713827bd09bSSatish Balay 		      used[i]=edge;
714827bd09bSSatish Balay 		    }
715827bd09bSSatish Balay 		}
716827bd09bSSatish Balay 	      if (ct>1) {ivec_sort(iptr,ct);}
717827bd09bSSatish Balay 
718827bd09bSSatish Balay 	      lnsep[edge]=ct;
71952f87cdaSBarry Smith 	      nsep[edge]= (PetscInt) rsum[1];
720827bd09bSSatish Balay 	      dir [edge]=RIGHT;
721827bd09bSSatish Balay 	    }
722827bd09bSSatish Balay 
723827bd09bSSatish Balay 	  /* LATER or we can recur on these to order seps at this level */
724827bd09bSSatish Balay 	  /* do we need full set of separators for this?                */
725827bd09bSSatish Balay 
726827bd09bSSatish Balay 	  /* fold rhs hc into lower */
727827bd09bSSatish Balay 	  if (id>=mask)
728827bd09bSSatish Balay 	    {id-=mask;}
729827bd09bSSatish Balay 	}
730827bd09bSSatish Balay     }
731827bd09bSSatish Balay   else
732827bd09bSSatish Balay     {
733827bd09bSSatish Balay       for (iptr=fo+n,id=my_id,mask=num_nodes>>1,edge=level;edge>0;edge--,mask>>=1)
734827bd09bSSatish Balay 	{
735827bd09bSSatish Balay 	  /* set rsh of hc, fire, and collect lhs responses */
736827bd09bSSatish Balay 	  (id<mask) ? rvec_zero(lhs,m) : rvec_set(lhs,1.0,m);
737827bd09bSSatish Balay 	  gs_gop_hc(gs_handle,lhs,"+\0",edge);
738827bd09bSSatish Balay 
739827bd09bSSatish Balay 	  /* set lsh of hc, fire, and collect rhs responses */
740827bd09bSSatish Balay 	  (id<mask) ? rvec_set(rhs,1.0,m) : rvec_zero(rhs,m);
741827bd09bSSatish Balay 	  gs_gop_hc(gs_handle,rhs,"+\0",edge);
742827bd09bSSatish Balay 
743827bd09bSSatish Balay 	  /* count number of dofs I own that have signal and not in sep set */
744827bd09bSSatish Balay 	  for (ivec_zero(sum,4),ct=i=0;i<n;i++)
745827bd09bSSatish Balay 	    {
746827bd09bSSatish Balay 	      if (!used[i])
747827bd09bSSatish Balay 		{
748827bd09bSSatish Balay 		  /* number of unmarked dofs on node */
749827bd09bSSatish Balay 		  ct++;
750827bd09bSSatish Balay 		  /* number of dofs to be marked on lhs hc */
751827bd09bSSatish Balay 		  if ((id< mask)&&(lhs[i]!=0.0)) {sum[0]++;}
752827bd09bSSatish Balay 		  /* number of dofs to be marked on rhs hc */
753827bd09bSSatish Balay 		  if ((id>=mask)&&(rhs[i]!=0.0)) {sum[1]++;}
754827bd09bSSatish Balay 		}
755827bd09bSSatish Balay 	    }
756827bd09bSSatish Balay 
757827bd09bSSatish Balay 	  /* go for load balance - choose half with most unmarked dofs, bias LHS */
758827bd09bSSatish Balay 	  (id<mask) ? (sum[2]=ct) : (sum[3]=ct);
759827bd09bSSatish Balay 	  giop_hc(sum,w,4,op,edge);
760827bd09bSSatish Balay 
761827bd09bSSatish Balay 	  /* lhs hc wins */
762827bd09bSSatish Balay 	  if (sum[2]>=sum[3])
763827bd09bSSatish Balay 	    {
764827bd09bSSatish Balay 	      if (id<mask)
765827bd09bSSatish Balay 		{
766827bd09bSSatish Balay 		  /* mark dofs I own that have signal and not in sep set */
767827bd09bSSatish Balay 		  for (ct=i=0;i<n;i++)
768827bd09bSSatish Balay 		    {
769827bd09bSSatish Balay 		      if ((!used[i])&&(lhs[i]!=0.0))
770827bd09bSSatish Balay 			{
771827bd09bSSatish Balay 			  ct++; nfo++;
772827bd09bSSatish Balay 			  *--iptr = local2global[i];
773827bd09bSSatish Balay 			  used[i]=edge;
774827bd09bSSatish Balay 			}
775827bd09bSSatish Balay 		    }
776827bd09bSSatish Balay 		  if (ct>1) {ivec_sort(iptr,ct);}
777827bd09bSSatish Balay 		  lnsep[edge]=ct;
778827bd09bSSatish Balay 		}
779827bd09bSSatish Balay 	      nsep[edge]=sum[0];
780827bd09bSSatish Balay 	      dir [edge]=LEFT;
781827bd09bSSatish Balay 	    }
782827bd09bSSatish Balay 	  /* rhs hc wins */
783827bd09bSSatish Balay 	  else
784827bd09bSSatish Balay 	    {
785827bd09bSSatish Balay 	      if (id>=mask)
786827bd09bSSatish Balay 		{
787827bd09bSSatish Balay 		  /* mark dofs I own that have signal and not in sep set */
788827bd09bSSatish Balay 		  for (ct=i=0;i<n;i++)
789827bd09bSSatish Balay 		    {
790827bd09bSSatish Balay 		      if ((!used[i])&&(rhs[i]!=0.0))
791827bd09bSSatish Balay 			{
792827bd09bSSatish Balay 			  ct++; nfo++;
793827bd09bSSatish Balay 			  *--iptr = local2global[i];
794827bd09bSSatish Balay 			  used[i]=edge;
795827bd09bSSatish Balay 			}
796827bd09bSSatish Balay 		    }
797827bd09bSSatish Balay 		  if (ct>1) {ivec_sort(iptr,ct);}
798827bd09bSSatish Balay 		  lnsep[edge]=ct;
799827bd09bSSatish Balay 		}
800827bd09bSSatish Balay 	      nsep[edge]=sum[1];
801827bd09bSSatish Balay 	      dir [edge]=RIGHT;
802827bd09bSSatish Balay 	    }
803827bd09bSSatish Balay 	  /* LATER or we can recur on these to order seps at this level */
804827bd09bSSatish Balay 	  /* do we need full set of separators for this?                */
805827bd09bSSatish Balay 
806827bd09bSSatish Balay 	  /* fold rhs hc into lower */
807827bd09bSSatish Balay 	  if (id>=mask)
808827bd09bSSatish Balay 	    {id-=mask;}
809827bd09bSSatish Balay 	}
810827bd09bSSatish Balay     }
811827bd09bSSatish Balay 
812827bd09bSSatish Balay 
813827bd09bSSatish Balay   /* level 0 is on processor case - so mark the remainder */
814827bd09bSSatish Balay   for (ct=i=0;i<n;i++)
815827bd09bSSatish Balay     {
816827bd09bSSatish Balay       if (!used[i])
817827bd09bSSatish Balay 	{
818827bd09bSSatish Balay 	  ct++; nfo++;
819827bd09bSSatish Balay 	  *--iptr = local2global[i];
820827bd09bSSatish Balay 	  used[i]=edge;
821827bd09bSSatish Balay 	}
822827bd09bSSatish Balay     }
823827bd09bSSatish Balay   if (ct>1) {ivec_sort(iptr,ct);}
824827bd09bSSatish Balay   lnsep[edge]=ct;
825827bd09bSSatish Balay   nsep [edge]=ct;
826827bd09bSSatish Balay   dir  [edge]=LEFT;
827827bd09bSSatish Balay 
828827bd09bSSatish Balay   xxt_handle->info->nsep=nsep;
829827bd09bSSatish Balay   xxt_handle->info->lnsep=lnsep;
830827bd09bSSatish Balay   xxt_handle->info->fo=fo;
831827bd09bSSatish Balay   xxt_handle->info->nfo=nfo;
832827bd09bSSatish Balay 
833a501084fSBarry Smith   free(dir);
834a501084fSBarry Smith   free(lhs);
835a501084fSBarry Smith   free(rhs);
836a501084fSBarry Smith   free(used);
8370924e98cSBarry Smith   PetscFunctionReturn(0);
838827bd09bSSatish Balay }
839827bd09bSSatish Balay 
8407b1ae94cSBarry Smith /**************************************xxt.c***********************************/
84152f87cdaSBarry Smith static mv_info *set_mvi(PetscInt *local2global, PetscInt n, PetscInt m, void *matvec, void *grid_data)
842827bd09bSSatish Balay {
843827bd09bSSatish Balay   mv_info *mvi;
844827bd09bSSatish Balay 
845827bd09bSSatish Balay 
846a501084fSBarry Smith   mvi = (mv_info*)malloc(sizeof(mv_info));
847827bd09bSSatish Balay   mvi->n=n;
848827bd09bSSatish Balay   mvi->m=m;
849827bd09bSSatish Balay   mvi->n_global=-1;
850827bd09bSSatish Balay   mvi->m_global=-1;
85152f87cdaSBarry Smith   mvi->local2global=(PetscInt*)malloc((m+1)*sizeof(PetscInt));
852827bd09bSSatish Balay   ivec_copy(mvi->local2global,local2global,m);
853827bd09bSSatish Balay   mvi->local2global[m] = INT_MAX;
854a501084fSBarry Smith   mvi->matvec=(PetscErrorCode (*)(mv_info*,PetscScalar*,PetscScalar*))matvec;
855827bd09bSSatish Balay   mvi->grid_data=grid_data;
856827bd09bSSatish Balay 
857827bd09bSSatish Balay   /* set xxt communication handle to perform restricted matvec */
858827bd09bSSatish Balay   mvi->gs_handle = gs_init(local2global, m, num_nodes);
859827bd09bSSatish Balay 
860827bd09bSSatish Balay   return(mvi);
861827bd09bSSatish Balay }
862827bd09bSSatish Balay 
8637b1ae94cSBarry Smith /**************************************xxt.c***********************************/
8640924e98cSBarry Smith static PetscErrorCode do_matvec(mv_info *A, PetscScalar *v, PetscScalar *u)
865827bd09bSSatish Balay {
8660924e98cSBarry Smith   PetscFunctionBegin;
867827bd09bSSatish Balay   A->matvec((mv_info*)A->grid_data,v,u);
8680924e98cSBarry Smith   PetscFunctionReturn(0);
869827bd09bSSatish Balay }
870827bd09bSSatish Balay 
871827bd09bSSatish Balay 
872827bd09bSSatish Balay 
873