xref: /petsc/src/ksp/pc/impls/tfs/xxt.c (revision a501084f0ec6f311bc0b878b7b6f8a6508165e40)
1827bd09bSSatish Balay 
2827bd09bSSatish Balay /*************************************xxt.c************************************
3827bd09bSSatish Balay Module Name: xxt
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 **************************************xxt.c***********************************/
19827bd09bSSatish Balay 
20827bd09bSSatish Balay 
21827bd09bSSatish Balay /*************************************xxt.c************************************
22827bd09bSSatish Balay NOTES ON USAGE:
23827bd09bSSatish Balay 
24827bd09bSSatish Balay **************************************xxt.c***********************************/
25827bd09bSSatish Balay #include <stdio.h>
26827bd09bSSatish Balay #include <stdlib.h>
27827bd09bSSatish Balay #include <limits.h>
28827bd09bSSatish Balay #include <float.h>
29827bd09bSSatish Balay #include <math.h>
3039945688SSatish Balay 
3139945688SSatish Balay #include "petsc.h"
32827bd09bSSatish Balay 
33827bd09bSSatish Balay #include "const.h"
34827bd09bSSatish Balay #include "types.h"
35827bd09bSSatish Balay #include "comm.h"
36827bd09bSSatish Balay #include "error.h"
37827bd09bSSatish Balay #include "ivec.h"
38827bd09bSSatish Balay #include "queue.h"
39827bd09bSSatish Balay #include "gs.h"
40827bd09bSSatish Balay #include "xxt.h"
41827bd09bSSatish Balay 
42827bd09bSSatish Balay #define LEFT  -1
43827bd09bSSatish Balay #define RIGHT  1
44827bd09bSSatish Balay #define BOTH   0
45827bd09bSSatish Balay #define MAX_FORTRAN_HANDLES  10
46827bd09bSSatish Balay 
47827bd09bSSatish Balay typedef struct xxt_solver_info {
48827bd09bSSatish Balay   int n, m, n_global, m_global;
49827bd09bSSatish Balay   int nnz, max_nnz, msg_buf_sz;
50827bd09bSSatish Balay   int *nsep, *lnsep, *fo, nfo, *stages;
51827bd09bSSatish Balay   int *col_sz, *col_indices;
52*a501084fSBarry Smith   PetscScalar **col_vals, *x, *solve_uu, *solve_w;
53827bd09bSSatish Balay   int nsolves;
54*a501084fSBarry Smith   PetscScalar tot_solve_time;
55827bd09bSSatish Balay } xxt_info;
56827bd09bSSatish Balay 
57827bd09bSSatish Balay typedef struct matvec_info {
58827bd09bSSatish Balay   int n, m, n_global, m_global;
59827bd09bSSatish Balay   int *local2global;
60827bd09bSSatish Balay   gs_ADT gs_handle;
61*a501084fSBarry Smith   PetscErrorCode (*matvec)(struct matvec_info*,PetscScalar*,PetscScalar*);
62827bd09bSSatish Balay   void *grid_data;
63827bd09bSSatish Balay } mv_info;
64827bd09bSSatish Balay 
65827bd09bSSatish Balay struct xxt_CDT{
66827bd09bSSatish Balay   int id;
67827bd09bSSatish Balay   int ns;
68827bd09bSSatish Balay   int level;
69827bd09bSSatish Balay   xxt_info *info;
70827bd09bSSatish Balay   mv_info  *mvi;
71827bd09bSSatish Balay };
72827bd09bSSatish Balay 
73827bd09bSSatish Balay static int n_xxt=0;
74827bd09bSSatish Balay static int n_xxt_handles=0;
75827bd09bSSatish Balay 
76827bd09bSSatish Balay /* prototypes */
77*a501084fSBarry Smith static void do_xxt_solve(xxt_ADT xxt_handle, PetscScalar *rhs);
78827bd09bSSatish Balay static void check_init(void);
79827bd09bSSatish Balay static void check_handle(xxt_ADT xxt_handle);
80827bd09bSSatish Balay static void det_separators(xxt_ADT xxt_handle);
81*a501084fSBarry Smith static void do_matvec(mv_info *A, PetscScalar *v, PetscScalar *u);
82827bd09bSSatish Balay static int xxt_generate(xxt_ADT xxt_handle);
83827bd09bSSatish Balay static int do_xxt_factor(xxt_ADT xxt_handle);
84827bd09bSSatish Balay static mv_info *set_mvi(int *local2global, int n, int m, void *matvec, void *grid_data);
85*a501084fSBarry Smith 
86827bd09bSSatish Balay 
87827bd09bSSatish Balay 
88827bd09bSSatish Balay /*************************************xxt.c************************************
89827bd09bSSatish Balay Function: XXT_new()
90827bd09bSSatish Balay 
91827bd09bSSatish Balay Input :
92827bd09bSSatish Balay Output:
93827bd09bSSatish Balay Return:
94827bd09bSSatish Balay Description:
95827bd09bSSatish Balay **************************************xxt.c***********************************/
96827bd09bSSatish Balay xxt_ADT
97827bd09bSSatish Balay XXT_new(void)
98827bd09bSSatish Balay {
99827bd09bSSatish Balay   xxt_ADT xxt_handle;
100827bd09bSSatish Balay 
101827bd09bSSatish Balay 
102827bd09bSSatish Balay 
103827bd09bSSatish Balay   /* rolling count on n_xxt ... pot. problem here */
104827bd09bSSatish Balay   n_xxt_handles++;
105*a501084fSBarry Smith   xxt_handle       = (xxt_ADT)malloc(sizeof(struct xxt_CDT));
106827bd09bSSatish Balay   xxt_handle->id   = ++n_xxt;
107827bd09bSSatish Balay   xxt_handle->info = NULL; xxt_handle->mvi  = NULL;
108827bd09bSSatish Balay 
109827bd09bSSatish Balay   return(xxt_handle);
110827bd09bSSatish Balay }
111827bd09bSSatish Balay 
112827bd09bSSatish Balay 
113827bd09bSSatish Balay /*************************************xxt.c************************************
114827bd09bSSatish Balay Function: XXT_factor()
115827bd09bSSatish Balay 
116827bd09bSSatish Balay Input :
117827bd09bSSatish Balay Output:
118827bd09bSSatish Balay Return:
119827bd09bSSatish Balay Description:
120827bd09bSSatish Balay **************************************xxt.c***********************************/
121827bd09bSSatish Balay int
122827bd09bSSatish Balay XXT_factor(xxt_ADT xxt_handle, /* prev. allocated xxt  handle */
123827bd09bSSatish Balay 	   int *local2global,  /* global column mapping       */
124827bd09bSSatish Balay 	   int n,              /* local num rows              */
125827bd09bSSatish Balay 	   int m,              /* local num cols              */
126827bd09bSSatish Balay 	   void *matvec,       /* b_loc=A_local.x_loc         */
127827bd09bSSatish Balay 	   void *grid_data     /* grid data for matvec        */
128827bd09bSSatish Balay 	   )
129827bd09bSSatish Balay {
130827bd09bSSatish Balay   check_init();
131827bd09bSSatish Balay   check_handle(xxt_handle);
132827bd09bSSatish Balay 
133827bd09bSSatish Balay   /* only 2^k for now and all nodes participating */
134827bd09bSSatish Balay   if ((1<<(xxt_handle->level=i_log2_num_nodes))!=num_nodes)
135827bd09bSSatish Balay     {error_msg_fatal("only 2^k for now and MPI_COMM_WORLD!!! %d != %d\n",1<<i_log2_num_nodes,num_nodes);}
136827bd09bSSatish Balay 
137827bd09bSSatish Balay   /* space for X info */
138*a501084fSBarry Smith   xxt_handle->info = (xxt_info*)malloc(sizeof(xxt_info));
139827bd09bSSatish Balay 
140827bd09bSSatish Balay   /* set up matvec handles */
141827bd09bSSatish Balay   xxt_handle->mvi  = set_mvi(local2global, n, m, matvec, grid_data);
142827bd09bSSatish Balay 
143827bd09bSSatish Balay   /* matrix is assumed to be of full rank */
144827bd09bSSatish Balay   /* LATER we can reset to indicate rank def. */
145827bd09bSSatish Balay   xxt_handle->ns=0;
146827bd09bSSatish Balay 
147827bd09bSSatish Balay   /* determine separators and generate firing order - NB xxt info set here */
148827bd09bSSatish Balay   det_separators(xxt_handle);
149827bd09bSSatish Balay 
150827bd09bSSatish Balay   return(do_xxt_factor(xxt_handle));
151827bd09bSSatish Balay }
152827bd09bSSatish Balay 
153827bd09bSSatish Balay 
154827bd09bSSatish Balay /*************************************xxt.c************************************
155827bd09bSSatish Balay Function: XXT_solve
156827bd09bSSatish Balay 
157827bd09bSSatish Balay Input :
158827bd09bSSatish Balay Output:
159827bd09bSSatish Balay Return:
160827bd09bSSatish Balay Description:
161827bd09bSSatish Balay **************************************xxt.c***********************************/
162827bd09bSSatish Balay int
163827bd09bSSatish Balay XXT_solve(xxt_ADT xxt_handle, double *x, double *b)
164827bd09bSSatish Balay {
165827bd09bSSatish Balay 
166827bd09bSSatish Balay   check_init();
167827bd09bSSatish Balay   check_handle(xxt_handle);
168827bd09bSSatish Balay 
169827bd09bSSatish Balay   /* need to copy b into x? */
170827bd09bSSatish Balay   if (b)
171827bd09bSSatish Balay     {rvec_copy(x,b,xxt_handle->mvi->n);}
172827bd09bSSatish Balay   do_xxt_solve(xxt_handle,x);
173827bd09bSSatish Balay 
174827bd09bSSatish Balay   return(0);
175827bd09bSSatish Balay }
176827bd09bSSatish Balay 
177827bd09bSSatish Balay 
178827bd09bSSatish Balay /*************************************xxt.c************************************
179827bd09bSSatish Balay Function: XXT_free()
180827bd09bSSatish Balay 
181827bd09bSSatish Balay Input :
182827bd09bSSatish Balay Output:
183827bd09bSSatish Balay Return:
184827bd09bSSatish Balay Description:
185827bd09bSSatish Balay **************************************xxt.c***********************************/
186827bd09bSSatish Balay int
187827bd09bSSatish Balay XXT_free(xxt_ADT xxt_handle)
188827bd09bSSatish Balay {
189827bd09bSSatish Balay 
190827bd09bSSatish Balay   check_init();
191827bd09bSSatish Balay   check_handle(xxt_handle);
192827bd09bSSatish Balay   n_xxt_handles--;
193827bd09bSSatish Balay 
194*a501084fSBarry Smith   free(xxt_handle->info->nsep);
195*a501084fSBarry Smith   free(xxt_handle->info->lnsep);
196*a501084fSBarry Smith   free(xxt_handle->info->fo);
197*a501084fSBarry Smith   free(xxt_handle->info->stages);
198*a501084fSBarry Smith   free(xxt_handle->info->solve_uu);
199*a501084fSBarry Smith   free(xxt_handle->info->solve_w);
200*a501084fSBarry Smith   free(xxt_handle->info->x);
201*a501084fSBarry Smith   free(xxt_handle->info->col_vals);
202*a501084fSBarry Smith   free(xxt_handle->info->col_sz);
203*a501084fSBarry Smith   free(xxt_handle->info->col_indices);
204*a501084fSBarry Smith   free(xxt_handle->info);
205*a501084fSBarry Smith   free(xxt_handle->mvi->local2global);
206827bd09bSSatish Balay    gs_free(xxt_handle->mvi->gs_handle);
207*a501084fSBarry Smith   free(xxt_handle->mvi);
208*a501084fSBarry Smith   free(xxt_handle);
209827bd09bSSatish Balay 
210827bd09bSSatish Balay 
211827bd09bSSatish Balay 
212827bd09bSSatish Balay   /* if the check fails we nuke */
213*a501084fSBarry Smith   /* if NULL pointer passed to free we nuke */
214827bd09bSSatish Balay   /* if the calls to free fail that's not my problem */
215827bd09bSSatish Balay   return(0);
216827bd09bSSatish Balay }
217827bd09bSSatish Balay 
218827bd09bSSatish Balay 
219827bd09bSSatish Balay 
220827bd09bSSatish Balay /*************************************xxt.c************************************
221827bd09bSSatish Balay Function:
222827bd09bSSatish Balay 
223827bd09bSSatish Balay Input :
224827bd09bSSatish Balay Output:
225827bd09bSSatish Balay Return:
226827bd09bSSatish Balay Description:
227827bd09bSSatish Balay **************************************xxt.c***********************************/
228827bd09bSSatish Balay int
229827bd09bSSatish Balay XXT_stats(xxt_ADT xxt_handle)
230827bd09bSSatish Balay {
231827bd09bSSatish Balay   int  op[] = {NON_UNIFORM,GL_MIN,GL_MAX,GL_ADD,GL_MIN,GL_MAX,GL_ADD,GL_MIN,GL_MAX,GL_ADD};
232827bd09bSSatish Balay   int fop[] = {NON_UNIFORM,GL_MIN,GL_MAX,GL_ADD};
233827bd09bSSatish Balay   int   vals[9],  work[9];
234*a501084fSBarry Smith   PetscScalar fvals[3], fwork[3];
235827bd09bSSatish Balay 
236827bd09bSSatish Balay 
237827bd09bSSatish Balay 
238827bd09bSSatish Balay   check_init();
239827bd09bSSatish Balay   check_handle(xxt_handle);
240827bd09bSSatish Balay 
241827bd09bSSatish Balay   /* if factorization not done there are no stats */
242827bd09bSSatish Balay   if (!xxt_handle->info||!xxt_handle->mvi)
243827bd09bSSatish Balay     {
244827bd09bSSatish Balay       if (!my_id)
245827bd09bSSatish Balay 	{printf("XXT_stats() :: no stats available!\n");}
246827bd09bSSatish Balay       return 1;
247827bd09bSSatish Balay     }
248827bd09bSSatish Balay 
249827bd09bSSatish Balay   vals[0]=vals[1]=vals[2]=xxt_handle->info->nnz;
250827bd09bSSatish Balay   vals[3]=vals[4]=vals[5]=xxt_handle->mvi->n;
251827bd09bSSatish Balay   vals[6]=vals[7]=vals[8]=xxt_handle->info->msg_buf_sz;
252827bd09bSSatish Balay   giop(vals,work,sizeof(op)/sizeof(op[0])-1,op);
253827bd09bSSatish Balay 
254827bd09bSSatish Balay   fvals[0]=fvals[1]=fvals[2]
255827bd09bSSatish Balay     =xxt_handle->info->tot_solve_time/xxt_handle->info->nsolves++;
256827bd09bSSatish Balay   grop(fvals,fwork,sizeof(fop)/sizeof(fop[0])-1,fop);
257827bd09bSSatish Balay 
258827bd09bSSatish Balay   if (!my_id)
259827bd09bSSatish Balay     {
260827bd09bSSatish Balay       printf("%d :: min   xxt_nnz=%d\n",my_id,vals[0]);
261827bd09bSSatish Balay       printf("%d :: max   xxt_nnz=%d\n",my_id,vals[1]);
262827bd09bSSatish Balay       printf("%d :: avg   xxt_nnz=%g\n",my_id,1.0*vals[2]/num_nodes);
263827bd09bSSatish Balay       printf("%d :: tot   xxt_nnz=%d\n",my_id,vals[2]);
264827bd09bSSatish Balay       printf("%d :: xxt   C(2d)  =%g\n",my_id,vals[2]/(pow(1.0*vals[5],1.5)));
265827bd09bSSatish Balay       printf("%d :: xxt   C(3d)  =%g\n",my_id,vals[2]/(pow(1.0*vals[5],1.6667)));
266827bd09bSSatish Balay       printf("%d :: min   xxt_n  =%d\n",my_id,vals[3]);
267827bd09bSSatish Balay       printf("%d :: max   xxt_n  =%d\n",my_id,vals[4]);
268827bd09bSSatish Balay       printf("%d :: avg   xxt_n  =%g\n",my_id,1.0*vals[5]/num_nodes);
269827bd09bSSatish Balay       printf("%d :: tot   xxt_n  =%d\n",my_id,vals[5]);
270827bd09bSSatish Balay       printf("%d :: min   xxt_buf=%d\n",my_id,vals[6]);
271827bd09bSSatish Balay       printf("%d :: max   xxt_buf=%d\n",my_id,vals[7]);
272827bd09bSSatish Balay       printf("%d :: avg   xxt_buf=%g\n",my_id,1.0*vals[8]/num_nodes);
273827bd09bSSatish Balay       printf("%d :: min   xxt_slv=%g\n",my_id,fvals[0]);
274827bd09bSSatish Balay       printf("%d :: max   xxt_slv=%g\n",my_id,fvals[1]);
275827bd09bSSatish Balay       printf("%d :: avg   xxt_slv=%g\n",my_id,fvals[2]/num_nodes);
276827bd09bSSatish Balay     }
277827bd09bSSatish Balay 
278827bd09bSSatish Balay   return(0);
279827bd09bSSatish Balay }
280827bd09bSSatish Balay 
281827bd09bSSatish Balay 
282827bd09bSSatish Balay /*************************************xxt.c************************************
283827bd09bSSatish Balay Function: do_xxt_factor
284827bd09bSSatish Balay 
285827bd09bSSatish Balay Input :
286827bd09bSSatish Balay Output:
287827bd09bSSatish Balay Return:
288827bd09bSSatish Balay Description: get A_local, local portion of global coarse matrix which
289827bd09bSSatish Balay is a row dist. nxm matrix w/ n<m.
290827bd09bSSatish Balay    o my_ml holds address of ML struct associated w/A_local and coarse grid
291827bd09bSSatish Balay    o local2global holds global number of column i (i=0,...,m-1)
292827bd09bSSatish Balay    o local2global holds global number of row    i (i=0,...,n-1)
293827bd09bSSatish Balay    o mylocmatvec performs A_local . vec_local (note that gs is performed using
294827bd09bSSatish Balay    gs_init/gop).
295827bd09bSSatish Balay 
296827bd09bSSatish Balay mylocmatvec = my_ml->Amat[grid_tag].matvec->external;
297827bd09bSSatish Balay mylocmatvec (void :: void *data, double *in, double *out)
298827bd09bSSatish Balay **************************************xxt.c***********************************/
299827bd09bSSatish Balay static
300827bd09bSSatish Balay int
301827bd09bSSatish Balay do_xxt_factor(xxt_ADT xxt_handle)
302827bd09bSSatish Balay {
303827bd09bSSatish Balay   int flag;
304827bd09bSSatish Balay 
305827bd09bSSatish Balay 
306827bd09bSSatish Balay   flag=xxt_generate(xxt_handle);
307827bd09bSSatish Balay 
308827bd09bSSatish Balay   return(flag);
309827bd09bSSatish Balay }
310827bd09bSSatish Balay 
311827bd09bSSatish Balay 
312827bd09bSSatish Balay /*************************************xxt.c************************************
313827bd09bSSatish Balay Function:
314827bd09bSSatish Balay 
315827bd09bSSatish Balay Input :
316827bd09bSSatish Balay Output:
317827bd09bSSatish Balay Return:
318827bd09bSSatish Balay Description:
319827bd09bSSatish Balay **************************************xxt.c***********************************/
320827bd09bSSatish Balay static
321827bd09bSSatish Balay int
322827bd09bSSatish Balay xxt_generate(xxt_ADT xxt_handle)
323827bd09bSSatish Balay {
324827bd09bSSatish Balay   int i,j,k,idex;
325827bd09bSSatish Balay   int dim, col;
326*a501084fSBarry Smith   PetscScalar *u, *uu, *v, *z, *w, alpha, alpha_w;
327827bd09bSSatish Balay   int *segs;
328827bd09bSSatish Balay   int op[] = {GL_ADD,0};
329827bd09bSSatish Balay   int off, len;
330*a501084fSBarry Smith   PetscScalar *x_ptr;
331827bd09bSSatish Balay   int *iptr, flag;
332827bd09bSSatish Balay   int start=0, end, work;
333827bd09bSSatish Balay   int op2[] = {GL_MIN,0};
334827bd09bSSatish Balay   gs_ADT gs_handle;
335827bd09bSSatish Balay   int *nsep, *lnsep, *fo;
336827bd09bSSatish Balay   int a_n=xxt_handle->mvi->n;
337827bd09bSSatish Balay   int a_m=xxt_handle->mvi->m;
338827bd09bSSatish Balay   int *a_local2global=xxt_handle->mvi->local2global;
339827bd09bSSatish Balay   int level;
340827bd09bSSatish Balay   int xxt_nnz=0, xxt_max_nnz=0;
341827bd09bSSatish Balay   int n, m;
342827bd09bSSatish Balay   int *col_sz, *col_indices, *stages;
343*a501084fSBarry Smith   PetscScalar **col_vals, *x;
344827bd09bSSatish Balay   int n_global;
345827bd09bSSatish Balay   int xxt_zero_nnz=0;
346827bd09bSSatish Balay   int xxt_zero_nnz_0=0;
347*a501084fSBarry Smith   PetscBLASInt i1 = 1;
348*a501084fSBarry Smith   PetscScalar dm1 = -1.0;
349827bd09bSSatish Balay 
350827bd09bSSatish Balay   n=xxt_handle->mvi->n;
351827bd09bSSatish Balay   nsep=xxt_handle->info->nsep;
352827bd09bSSatish Balay   lnsep=xxt_handle->info->lnsep;
353827bd09bSSatish Balay   fo=xxt_handle->info->fo;
354827bd09bSSatish Balay   end=lnsep[0];
355827bd09bSSatish Balay   level=xxt_handle->level;
356827bd09bSSatish Balay   gs_handle=xxt_handle->mvi->gs_handle;
357827bd09bSSatish Balay 
358827bd09bSSatish Balay   /* is there a null space? */
359827bd09bSSatish Balay   /* LATER add in ability to detect null space by checking alpha */
360827bd09bSSatish Balay   for (i=0, j=0; i<=level; i++)
361827bd09bSSatish Balay     {j+=nsep[i];}
362827bd09bSSatish Balay 
363827bd09bSSatish Balay   m = j-xxt_handle->ns;
364827bd09bSSatish Balay   if (m!=j)
365827bd09bSSatish Balay     {printf("xxt_generate() :: null space exists %d %d %d\n",m,j,xxt_handle->ns);}
366827bd09bSSatish Balay 
367827bd09bSSatish Balay   /* get and initialize storage for x local         */
368827bd09bSSatish Balay   /* note that x local is nxm and stored by columns */
369*a501084fSBarry Smith   col_sz = (int*) malloc(m*sizeof(PetscInt));
370*a501084fSBarry Smith   col_indices = (int*) malloc((2*m+1)*sizeof(int));
371*a501084fSBarry Smith   col_vals = (PetscScalar **) malloc(m*sizeof(PetscScalar *));
372827bd09bSSatish Balay   for (i=j=0; i<m; i++, j+=2)
373827bd09bSSatish Balay     {
374827bd09bSSatish Balay       col_indices[j]=col_indices[j+1]=col_sz[i]=-1;
375827bd09bSSatish Balay       col_vals[i] = NULL;
376827bd09bSSatish Balay     }
377827bd09bSSatish Balay   col_indices[j]=-1;
378827bd09bSSatish Balay 
379827bd09bSSatish Balay   /* size of separators for each sub-hc working from bottom of tree to top */
380827bd09bSSatish Balay   /* this looks like nsep[]=segments */
381*a501084fSBarry Smith   stages = (int*) malloc((level+1)*sizeof(PetscInt));
382*a501084fSBarry Smith   segs   = (int*) malloc((level+1)*sizeof(PetscInt));
383827bd09bSSatish Balay   ivec_zero(stages,level+1);
384827bd09bSSatish Balay   ivec_copy(segs,nsep,level+1);
385827bd09bSSatish Balay   for (i=0; i<level; i++)
386827bd09bSSatish Balay     {segs[i+1] += segs[i];}
387827bd09bSSatish Balay   stages[0] = segs[0];
388827bd09bSSatish Balay 
389827bd09bSSatish Balay   /* temporary vectors  */
390*a501084fSBarry Smith   u  = (PetscScalar *) malloc(n*sizeof(PetscScalar));
391*a501084fSBarry Smith   z  = (PetscScalar *) malloc(n*sizeof(PetscScalar));
392*a501084fSBarry Smith   v  = (PetscScalar *) malloc(a_m*sizeof(PetscScalar));
393*a501084fSBarry Smith   uu = (PetscScalar *) malloc(m*sizeof(PetscScalar));
394*a501084fSBarry Smith   w  = (PetscScalar *) malloc(m*sizeof(PetscScalar));
395827bd09bSSatish Balay 
396827bd09bSSatish Balay   /* extra nnz due to replication of vertices across separators */
397827bd09bSSatish Balay   for (i=1, j=0; i<=level; i++)
398827bd09bSSatish Balay     {j+=nsep[i];}
399827bd09bSSatish Balay 
400827bd09bSSatish Balay   /* storage for sparse x values */
401827bd09bSSatish Balay   n_global = xxt_handle->info->n_global;
402827bd09bSSatish Balay   xxt_max_nnz = (int)(2.5*pow(1.0*n_global,1.6667) + j*n/2)/num_nodes;
403*a501084fSBarry Smith   x = (PetscScalar *) malloc(xxt_max_nnz*sizeof(PetscScalar));
404827bd09bSSatish Balay   xxt_nnz = 0;
405827bd09bSSatish Balay 
406827bd09bSSatish Balay   /* LATER - can embed next sep to fire in gs */
407827bd09bSSatish Balay   /* time to make the donuts - generate X factor */
408827bd09bSSatish Balay   for (dim=i=j=0;i<m;i++)
409827bd09bSSatish Balay     {
410827bd09bSSatish Balay       /* time to move to the next level? */
411827bd09bSSatish Balay       while (i==segs[dim])
412827bd09bSSatish Balay 	{
413827bd09bSSatish Balay #ifdef SAFE
414827bd09bSSatish Balay 	  if (dim==level)
415827bd09bSSatish Balay 	    {error_msg_fatal("dim about to exceed level\n"); break;}
416827bd09bSSatish Balay #endif
417827bd09bSSatish Balay 
418827bd09bSSatish Balay 	  stages[dim++]=i;
419827bd09bSSatish Balay 	  end+=lnsep[dim];
420827bd09bSSatish Balay 	}
421827bd09bSSatish Balay       stages[dim]=i;
422827bd09bSSatish Balay 
423827bd09bSSatish Balay       /* which column are we firing? */
424827bd09bSSatish Balay       /* i.e. set v_l */
425827bd09bSSatish Balay       /* use new seps and do global min across hc to determine which one to fire */
426827bd09bSSatish Balay       (start<end) ? (col=fo[start]) : (col=INT_MAX);
427827bd09bSSatish Balay       giop_hc(&col,&work,1,op2,dim);
428827bd09bSSatish Balay 
429827bd09bSSatish Balay       /* shouldn't need this */
430827bd09bSSatish Balay       if (col==INT_MAX)
431827bd09bSSatish Balay 	{
432827bd09bSSatish Balay 	  error_msg_warning("hey ... col==INT_MAX??\n");
433827bd09bSSatish Balay 	  continue;
434827bd09bSSatish Balay 	}
435827bd09bSSatish Balay 
436827bd09bSSatish Balay       /* do I own it? I should */
437827bd09bSSatish Balay       rvec_zero(v ,a_m);
438827bd09bSSatish Balay       if (col==fo[start])
439827bd09bSSatish Balay 	{
440827bd09bSSatish Balay 	  start++;
441827bd09bSSatish Balay 	  idex=ivec_linear_search(col, a_local2global, a_n);
442827bd09bSSatish Balay 	  if (idex!=-1)
443827bd09bSSatish Balay 	    {v[idex] = 1.0; j++;}
444827bd09bSSatish Balay 	  else
445827bd09bSSatish Balay 	    {error_msg_fatal("NOT FOUND!\n");}
446827bd09bSSatish Balay 	}
447827bd09bSSatish Balay       else
448827bd09bSSatish Balay 	{
449827bd09bSSatish Balay 	  idex=ivec_linear_search(col, a_local2global, a_m);
450827bd09bSSatish Balay 	  if (idex!=-1)
451827bd09bSSatish Balay 	    {v[idex] = 1.0;}
452827bd09bSSatish Balay 	}
453827bd09bSSatish Balay 
454827bd09bSSatish Balay       /* perform u = A.v_l */
455827bd09bSSatish Balay       rvec_zero(u,n);
456827bd09bSSatish Balay       do_matvec(xxt_handle->mvi,v,u);
457827bd09bSSatish Balay 
458827bd09bSSatish Balay       /* uu =  X^T.u_l (local portion) */
459827bd09bSSatish Balay       /* technically only need to zero out first i entries */
460827bd09bSSatish Balay       /* later turn this into an XXT_solve call ? */
461827bd09bSSatish Balay       rvec_zero(uu,m);
462827bd09bSSatish Balay       x_ptr=x;
463827bd09bSSatish Balay       iptr = col_indices;
464827bd09bSSatish Balay       for (k=0; k<i; k++)
465827bd09bSSatish Balay 	{
466827bd09bSSatish Balay 	  off = *iptr++;
467827bd09bSSatish Balay 	  len = *iptr++;
468827bd09bSSatish Balay 
469*a501084fSBarry Smith 	  uu[k] = BLdot_(&len,u+off,&i1,x_ptr,&i1);
470827bd09bSSatish Balay 	  x_ptr+=len;
471827bd09bSSatish Balay 	}
472827bd09bSSatish Balay 
473827bd09bSSatish Balay 
474827bd09bSSatish Balay       /* uu = X^T.u_l (comm portion) */
475827bd09bSSatish Balay       ssgl_radd  (uu, w, dim, stages);
476827bd09bSSatish Balay 
477827bd09bSSatish Balay       /* z = X.uu */
478827bd09bSSatish Balay       rvec_zero(z,n);
479827bd09bSSatish Balay       x_ptr=x;
480827bd09bSSatish Balay       iptr = col_indices;
481827bd09bSSatish Balay       for (k=0; k<i; k++)
482827bd09bSSatish Balay 	{
483827bd09bSSatish Balay 	  off = *iptr++;
484827bd09bSSatish Balay 	  len = *iptr++;
485827bd09bSSatish Balay 
486*a501084fSBarry Smith 	  BLaxpy_(&len,&uu[k],x_ptr,&i1,z+off,&i1);
487827bd09bSSatish Balay 	  x_ptr+=len;
488827bd09bSSatish Balay 	}
489827bd09bSSatish Balay 
490827bd09bSSatish Balay       /* compute v_l = v_l - z */
491827bd09bSSatish Balay       rvec_zero(v+a_n,a_m-a_n);
492*a501084fSBarry Smith       BLaxpy_(&n,&dm1,z,&i1,v,&i1);
493827bd09bSSatish Balay 
494827bd09bSSatish Balay       /* compute u_l = A.v_l */
495827bd09bSSatish Balay       if (a_n!=a_m)
496827bd09bSSatish Balay 	{gs_gop_hc(gs_handle,v,"+\0",dim);}
497827bd09bSSatish Balay       rvec_zero(u,n);
498827bd09bSSatish Balay       do_matvec(xxt_handle->mvi,v,u);
499827bd09bSSatish Balay 
500827bd09bSSatish Balay       /* compute sqrt(alpha) = sqrt(v_l^T.u_l) - local portion */
501*a501084fSBarry Smith       alpha = BLdot_(&n,u,&i1,v,&i1);
502827bd09bSSatish Balay       /* compute sqrt(alpha) = sqrt(v_l^T.u_l) - comm portion */
503827bd09bSSatish Balay       grop_hc(&alpha, &alpha_w, 1, op, dim);
504827bd09bSSatish Balay 
505*a501084fSBarry Smith       alpha = (PetscScalar) sqrt((double)alpha);
506827bd09bSSatish Balay 
507827bd09bSSatish Balay       /* check for small alpha                             */
508827bd09bSSatish Balay       /* LATER use this to detect and determine null space */
509827bd09bSSatish Balay       if (fabs(alpha)<1.0e-14)
510827bd09bSSatish Balay 	{error_msg_fatal("bad alpha! %g\n",alpha);}
511827bd09bSSatish Balay 
512827bd09bSSatish Balay       /* compute v_l = v_l/sqrt(alpha) */
513827bd09bSSatish Balay       rvec_scale(v,1.0/alpha,n);
514827bd09bSSatish Balay 
515827bd09bSSatish Balay       /* add newly generated column, v_l, to X */
516827bd09bSSatish Balay       flag = 1;
517827bd09bSSatish Balay       off=len=0;
518827bd09bSSatish Balay       for (k=0; k<n; k++)
519827bd09bSSatish Balay 	{
520827bd09bSSatish Balay 	  if (v[k]!=0.0)
521827bd09bSSatish Balay 	    {
522827bd09bSSatish Balay 	      len=k;
523827bd09bSSatish Balay 	      if (flag)
524827bd09bSSatish Balay 		{off=k; flag=0;}
525827bd09bSSatish Balay 	    }
526827bd09bSSatish Balay 	}
527827bd09bSSatish Balay 
528827bd09bSSatish Balay       len -= (off-1);
529827bd09bSSatish Balay 
530827bd09bSSatish Balay       if (len>0)
531827bd09bSSatish Balay 	{
532827bd09bSSatish Balay 	  if ((xxt_nnz+len)>xxt_max_nnz)
533827bd09bSSatish Balay 	    {
534827bd09bSSatish Balay 	      error_msg_warning("increasing space for X by 2x!\n");
535827bd09bSSatish Balay 	      xxt_max_nnz *= 2;
536*a501084fSBarry Smith 	      x_ptr = (PetscScalar *) malloc(xxt_max_nnz*sizeof(PetscScalar));
537827bd09bSSatish Balay 	      rvec_copy(x_ptr,x,xxt_nnz);
538*a501084fSBarry Smith 	      free(x);
539827bd09bSSatish Balay 	      x = x_ptr;
540827bd09bSSatish Balay 	      x_ptr+=xxt_nnz;
541827bd09bSSatish Balay 	    }
542827bd09bSSatish Balay 	  xxt_nnz += len;
543827bd09bSSatish Balay 	  rvec_copy(x_ptr,v+off,len);
544827bd09bSSatish Balay 
545827bd09bSSatish Balay           /* keep track of number of zeros */
546827bd09bSSatish Balay 	  if (dim)
547827bd09bSSatish Balay 	    {
548827bd09bSSatish Balay 	      for (k=0; k<len; k++)
549827bd09bSSatish Balay 		{
550827bd09bSSatish Balay 		  if (x_ptr[k]==0.0)
551827bd09bSSatish Balay 		    {xxt_zero_nnz++;}
552827bd09bSSatish Balay 		}
553827bd09bSSatish Balay 	    }
554827bd09bSSatish Balay 	  else
555827bd09bSSatish Balay 	    {
556827bd09bSSatish Balay 	      for (k=0; k<len; k++)
557827bd09bSSatish Balay 		{
558827bd09bSSatish Balay 		  if (x_ptr[k]==0.0)
559827bd09bSSatish Balay 		    {xxt_zero_nnz_0++;}
560827bd09bSSatish Balay 		}
561827bd09bSSatish Balay 	    }
562827bd09bSSatish Balay 	  col_indices[2*i] = off;
563827bd09bSSatish Balay 	  col_sz[i] = col_indices[2*i+1] = len;
564827bd09bSSatish Balay 	  col_vals[i] = x_ptr;
565827bd09bSSatish Balay 	}
566827bd09bSSatish Balay       else
567827bd09bSSatish Balay 	{
568827bd09bSSatish Balay 	  col_indices[2*i] = 0;
569827bd09bSSatish Balay 	  col_sz[i] = col_indices[2*i+1] = 0;
570827bd09bSSatish Balay 	  col_vals[i] = x_ptr;
571827bd09bSSatish Balay 	}
572827bd09bSSatish Balay     }
573827bd09bSSatish Balay 
574827bd09bSSatish Balay   /* close off stages for execution phase */
575827bd09bSSatish Balay   while (dim!=level)
576827bd09bSSatish Balay     {
577827bd09bSSatish Balay       stages[dim++]=i;
578827bd09bSSatish Balay       error_msg_warning("disconnected!!! dim(%d)!=level(%d)\n",dim,level);
579827bd09bSSatish Balay     }
580827bd09bSSatish Balay   stages[dim]=i;
581827bd09bSSatish Balay 
582827bd09bSSatish Balay   xxt_handle->info->n=xxt_handle->mvi->n;
583827bd09bSSatish Balay   xxt_handle->info->m=m;
584827bd09bSSatish Balay   xxt_handle->info->nnz=xxt_nnz;
585827bd09bSSatish Balay   xxt_handle->info->max_nnz=xxt_max_nnz;
586827bd09bSSatish Balay   xxt_handle->info->msg_buf_sz=stages[level]-stages[0];
587*a501084fSBarry Smith   xxt_handle->info->solve_uu = (PetscScalar *) malloc(m*sizeof(PetscScalar));
588*a501084fSBarry Smith   xxt_handle->info->solve_w  = (PetscScalar *) malloc(m*sizeof(PetscScalar));
589827bd09bSSatish Balay   xxt_handle->info->x=x;
590827bd09bSSatish Balay   xxt_handle->info->col_vals=col_vals;
591827bd09bSSatish Balay   xxt_handle->info->col_sz=col_sz;
592827bd09bSSatish Balay   xxt_handle->info->col_indices=col_indices;
593827bd09bSSatish Balay   xxt_handle->info->stages=stages;
594827bd09bSSatish Balay   xxt_handle->info->nsolves=0;
595827bd09bSSatish Balay   xxt_handle->info->tot_solve_time=0.0;
596827bd09bSSatish Balay 
597*a501084fSBarry Smith   free(segs);
598*a501084fSBarry Smith   free(u);
599*a501084fSBarry Smith   free(v);
600*a501084fSBarry Smith   free(uu);
601*a501084fSBarry Smith   free(z);
602*a501084fSBarry Smith   free(w);
603827bd09bSSatish Balay 
604827bd09bSSatish Balay   return(0);
605827bd09bSSatish Balay }
606827bd09bSSatish Balay 
607827bd09bSSatish Balay 
608827bd09bSSatish Balay /*************************************xxt.c************************************
609827bd09bSSatish Balay Function:
610827bd09bSSatish Balay 
611827bd09bSSatish Balay Input :
612827bd09bSSatish Balay Output:
613827bd09bSSatish Balay Return:
614827bd09bSSatish Balay Description:
615827bd09bSSatish Balay **************************************xxt.c***********************************/
616827bd09bSSatish Balay static
617827bd09bSSatish Balay void
618*a501084fSBarry Smith do_xxt_solve(xxt_ADT xxt_handle,  PetscScalar *uc)
619827bd09bSSatish Balay {
620*a501084fSBarry Smith    int off, len, *iptr;
621827bd09bSSatish Balay   int level       =xxt_handle->level;
622827bd09bSSatish Balay   int n           =xxt_handle->info->n;
623827bd09bSSatish Balay   int m           =xxt_handle->info->m;
624827bd09bSSatish Balay   int *stages     =xxt_handle->info->stages;
625827bd09bSSatish Balay   int *col_indices=xxt_handle->info->col_indices;
626*a501084fSBarry Smith   PetscScalar *x_ptr, *uu_ptr;
627*a501084fSBarry Smith   PetscScalar *solve_uu=xxt_handle->info->solve_uu;
628*a501084fSBarry Smith   PetscScalar *solve_w =xxt_handle->info->solve_w;
629*a501084fSBarry Smith   PetscScalar *x       =xxt_handle->info->x;
630*a501084fSBarry Smith   PetscBLASInt i1 = 1;
631827bd09bSSatish Balay 
632827bd09bSSatish Balay   uu_ptr=solve_uu;
633827bd09bSSatish Balay   rvec_zero(uu_ptr,m);
634827bd09bSSatish Balay 
635827bd09bSSatish Balay   /* x  = X.Y^T.b */
636827bd09bSSatish Balay   /* uu = Y^T.b */
637827bd09bSSatish Balay   for (x_ptr=x,iptr=col_indices; *iptr!=-1; x_ptr+=len)
638827bd09bSSatish Balay     {
639827bd09bSSatish Balay       off=*iptr++; len=*iptr++;
640*a501084fSBarry Smith       *uu_ptr++ = BLdot_(&len,uc+off,&i1,x_ptr,&i1);
641827bd09bSSatish Balay     }
642827bd09bSSatish Balay 
643827bd09bSSatish Balay   /* comunication of beta */
644827bd09bSSatish Balay   uu_ptr=solve_uu;
645827bd09bSSatish Balay   if (level) {ssgl_radd(uu_ptr, solve_w, level, stages);}
646827bd09bSSatish Balay 
647827bd09bSSatish Balay   rvec_zero(uc,n);
648827bd09bSSatish Balay 
649827bd09bSSatish Balay   /* x = X.uu */
650827bd09bSSatish Balay   for (x_ptr=x,iptr=col_indices; *iptr!=-1; x_ptr+=len)
651827bd09bSSatish Balay     {
652827bd09bSSatish Balay       off=*iptr++; len=*iptr++;
653*a501084fSBarry Smith       BLaxpy_(&len,uu_ptr++,x_ptr,&i1,uc+off,&i1);
654827bd09bSSatish Balay     }
655827bd09bSSatish Balay 
656827bd09bSSatish Balay }
657827bd09bSSatish Balay 
658827bd09bSSatish Balay 
659827bd09bSSatish Balay /*************************************Xxt.c************************************
660827bd09bSSatish Balay Function: check_init
661827bd09bSSatish Balay 
662827bd09bSSatish Balay Input :
663827bd09bSSatish Balay Output:
664827bd09bSSatish Balay Return:
665827bd09bSSatish Balay Description:
666827bd09bSSatish Balay **************************************xxt.c***********************************/
667827bd09bSSatish Balay static
668827bd09bSSatish Balay void
669827bd09bSSatish Balay check_init(void)
670827bd09bSSatish Balay {
671827bd09bSSatish Balay   comm_init();
672827bd09bSSatish Balay 
673827bd09bSSatish Balay }
674827bd09bSSatish Balay 
675827bd09bSSatish Balay 
676827bd09bSSatish Balay /*************************************xxt.c************************************
677827bd09bSSatish Balay Function: check_handle()
678827bd09bSSatish Balay 
679827bd09bSSatish Balay Input :
680827bd09bSSatish Balay Output:
681827bd09bSSatish Balay Return:
682827bd09bSSatish Balay Description:
683827bd09bSSatish Balay **************************************xxt.c***********************************/
684827bd09bSSatish Balay static
685827bd09bSSatish Balay void
686827bd09bSSatish Balay check_handle(xxt_ADT xxt_handle)
687827bd09bSSatish Balay {
688827bd09bSSatish Balay #ifdef SAFE
689827bd09bSSatish Balay   int vals[2], work[2], op[] = {NON_UNIFORM,GL_MIN,GL_MAX};
690827bd09bSSatish Balay #endif
691827bd09bSSatish Balay 
692827bd09bSSatish Balay 
693827bd09bSSatish Balay   if (xxt_handle==NULL)
694827bd09bSSatish Balay     {error_msg_fatal("check_handle() :: bad handle :: NULL %d\n",xxt_handle);}
695827bd09bSSatish Balay 
696827bd09bSSatish Balay #ifdef SAFE
697827bd09bSSatish Balay   vals[0]=vals[1]=xxt_handle->id;
698827bd09bSSatish Balay   giop(vals,work,sizeof(op)/sizeof(op[0])-1,op);
699827bd09bSSatish Balay   if ((vals[0]!=vals[1])||(xxt_handle->id<=0))
700827bd09bSSatish Balay     {error_msg_fatal("check_handle() :: bad handle :: id mismatch min/max %d/%d %d\n",
701827bd09bSSatish Balay 		     vals[0],vals[1], xxt_handle->id);}
702827bd09bSSatish Balay #endif
703827bd09bSSatish Balay 
704827bd09bSSatish Balay }
705827bd09bSSatish Balay 
706827bd09bSSatish Balay 
707827bd09bSSatish Balay /*************************************xxt.c************************************
708827bd09bSSatish Balay Function: det_separators
709827bd09bSSatish Balay 
710827bd09bSSatish Balay Input :
711827bd09bSSatish Balay Output:
712827bd09bSSatish Balay Return:
713827bd09bSSatish Balay Description:
714827bd09bSSatish Balay   det_separators(xxt_handle, local2global, n, m, mylocmatvec, grid_data);
715827bd09bSSatish Balay **************************************xxt.c***********************************/
716827bd09bSSatish Balay static
717827bd09bSSatish Balay void
718827bd09bSSatish Balay det_separators(xxt_ADT xxt_handle)
719827bd09bSSatish Balay {
720827bd09bSSatish Balay   int i, ct, id;
721827bd09bSSatish Balay   int mask, edge, *iptr;
722827bd09bSSatish Balay   int *dir, *used;
723827bd09bSSatish Balay   int sum[4], w[4];
724*a501084fSBarry Smith   PetscScalar rsum[4], rw[4];
725827bd09bSSatish Balay   int op[] = {GL_ADD,0};
726*a501084fSBarry Smith   PetscScalar *lhs, *rhs;
727827bd09bSSatish Balay   int *nsep, *lnsep, *fo, nfo=0;
728827bd09bSSatish Balay   gs_ADT gs_handle=xxt_handle->mvi->gs_handle;
729827bd09bSSatish Balay   int *local2global=xxt_handle->mvi->local2global;
730827bd09bSSatish Balay   int  n=xxt_handle->mvi->n;
731827bd09bSSatish Balay   int  m=xxt_handle->mvi->m;
732827bd09bSSatish Balay   int level=xxt_handle->level;
733827bd09bSSatish Balay   int shared=FALSE;
734827bd09bSSatish Balay 
735*a501084fSBarry Smith   dir  = (int*)malloc(sizeof(PetscInt)*(level+1));
736*a501084fSBarry Smith   nsep = (int*)malloc(sizeof(PetscInt)*(level+1));
737*a501084fSBarry Smith   lnsep= (int*)malloc(sizeof(PetscInt)*(level+1));
738*a501084fSBarry Smith   fo   = (int*)malloc(sizeof(PetscInt)*(n+1));
739*a501084fSBarry Smith   used = (int*)malloc(sizeof(PetscInt)*n);
740827bd09bSSatish Balay 
741827bd09bSSatish Balay   ivec_zero(dir  ,level+1);
742827bd09bSSatish Balay   ivec_zero(nsep ,level+1);
743827bd09bSSatish Balay   ivec_zero(lnsep,level+1);
744827bd09bSSatish Balay   ivec_set (fo   ,-1,n+1);
745827bd09bSSatish Balay   ivec_zero(used,n);
746827bd09bSSatish Balay 
747*a501084fSBarry Smith   lhs  = (double*)malloc(sizeof(PetscScalar)*m);
748*a501084fSBarry Smith   rhs  = (double*)malloc(sizeof(PetscScalar)*m);
749827bd09bSSatish Balay 
750827bd09bSSatish Balay   /* determine the # of unique dof */
751827bd09bSSatish Balay   rvec_zero(lhs,m);
752827bd09bSSatish Balay   rvec_set(lhs,1.0,n);
753827bd09bSSatish Balay   gs_gop_hc(gs_handle,lhs,"+\0",level);
754827bd09bSSatish Balay   rvec_zero(rsum,2);
755827bd09bSSatish Balay   for (ct=i=0;i<n;i++)
756827bd09bSSatish Balay     {
757827bd09bSSatish Balay       if (lhs[i]!=0.0)
758827bd09bSSatish Balay 	{rsum[0]+=1.0/lhs[i]; rsum[1]+=lhs[i];}
759827bd09bSSatish Balay     }
760827bd09bSSatish Balay   grop_hc(rsum,rw,2,op,level);
761827bd09bSSatish Balay   rsum[0]+=0.1;
762827bd09bSSatish Balay   rsum[1]+=0.1;
763827bd09bSSatish Balay   /*  if (!my_id)
764827bd09bSSatish Balay     {
765827bd09bSSatish Balay       printf("xxt n unique = %d (%g)\n",(int) rsum[0], rsum[0]);
766827bd09bSSatish Balay       printf("xxt n shared = %d (%g)\n",(int) rsum[1], rsum[1]);
767827bd09bSSatish Balay       }*/
768827bd09bSSatish Balay 
769827bd09bSSatish Balay   if (fabs(rsum[0]-rsum[1])>EPS)
770827bd09bSSatish Balay     {shared=TRUE;}
771827bd09bSSatish Balay 
772827bd09bSSatish Balay   xxt_handle->info->n_global=xxt_handle->info->m_global=(int) rsum[0];
773827bd09bSSatish Balay   xxt_handle->mvi->n_global =xxt_handle->mvi->m_global =(int) rsum[0];
774827bd09bSSatish Balay 
775827bd09bSSatish Balay   /* determine separator sets top down */
776827bd09bSSatish Balay   if (shared)
777827bd09bSSatish Balay     {
778827bd09bSSatish Balay       for (iptr=fo+n,id=my_id,mask=num_nodes>>1,edge=level;edge>0;edge--,mask>>=1)
779827bd09bSSatish Balay 	{
780827bd09bSSatish Balay 	  /* set rsh of hc, fire, and collect lhs responses */
781827bd09bSSatish Balay 	  (id<mask) ? rvec_zero(lhs,m) : rvec_set(lhs,1.0,m);
782827bd09bSSatish Balay 	  gs_gop_hc(gs_handle,lhs,"+\0",edge);
783827bd09bSSatish Balay 
784827bd09bSSatish Balay 	  /* set lsh of hc, fire, and collect rhs responses */
785827bd09bSSatish Balay 	  (id<mask) ? rvec_set(rhs,1.0,m) : rvec_zero(rhs,m);
786827bd09bSSatish Balay 	  gs_gop_hc(gs_handle,rhs,"+\0",edge);
787827bd09bSSatish Balay 
788827bd09bSSatish Balay 	  for (i=0;i<n;i++)
789827bd09bSSatish Balay 	    {
790827bd09bSSatish Balay 	      if (id< mask)
791827bd09bSSatish Balay 		{
792827bd09bSSatish Balay 		  if (lhs[i]!=0.0)
793827bd09bSSatish Balay 		    {lhs[i]=1.0;}
794827bd09bSSatish Balay 		}
795827bd09bSSatish Balay 	      if (id>=mask)
796827bd09bSSatish Balay 		{
797827bd09bSSatish Balay 		  if (rhs[i]!=0.0)
798827bd09bSSatish Balay 		    {rhs[i]=1.0;}
799827bd09bSSatish Balay 		}
800827bd09bSSatish Balay 	    }
801827bd09bSSatish Balay 
802827bd09bSSatish Balay 	  if (id< mask)
803827bd09bSSatish Balay 	    {gs_gop_hc(gs_handle,lhs,"+\0",edge-1);}
804827bd09bSSatish Balay 	  else
805827bd09bSSatish Balay 	    {gs_gop_hc(gs_handle,rhs,"+\0",edge-1);}
806827bd09bSSatish Balay 
807827bd09bSSatish Balay 	  /* count number of dofs I own that have signal and not in sep set */
808827bd09bSSatish Balay 	  rvec_zero(rsum,4);
809827bd09bSSatish Balay 	  for (ivec_zero(sum,4),ct=i=0;i<n;i++)
810827bd09bSSatish Balay 	    {
811827bd09bSSatish Balay 	      if (!used[i])
812827bd09bSSatish Balay 		{
813827bd09bSSatish Balay 		  /* number of unmarked dofs on node */
814827bd09bSSatish Balay 		  ct++;
815827bd09bSSatish Balay 		  /* number of dofs to be marked on lhs hc */
816827bd09bSSatish Balay 		  if (id< mask)
817827bd09bSSatish Balay 		    {
818827bd09bSSatish Balay 		      if (lhs[i]!=0.0)
819827bd09bSSatish Balay 			{sum[0]++; rsum[0]+=1.0/lhs[i];}
820827bd09bSSatish Balay 		    }
821827bd09bSSatish Balay 		  /* number of dofs to be marked on rhs hc */
822827bd09bSSatish Balay 		  if (id>=mask)
823827bd09bSSatish Balay 		    {
824827bd09bSSatish Balay 		      if (rhs[i]!=0.0)
825827bd09bSSatish Balay 			{sum[1]++; rsum[1]+=1.0/rhs[i];}
826827bd09bSSatish Balay 		    }
827827bd09bSSatish Balay 		}
828827bd09bSSatish Balay 	    }
829827bd09bSSatish Balay 
830827bd09bSSatish Balay 	  /* go for load balance - choose half with most unmarked dofs, bias LHS */
831827bd09bSSatish Balay 	  (id<mask) ? (sum[2]=ct) : (sum[3]=ct);
832827bd09bSSatish Balay 	  (id<mask) ? (rsum[2]=ct) : (rsum[3]=ct);
833827bd09bSSatish Balay 	  giop_hc(sum,w,4,op,edge);
834827bd09bSSatish Balay 	  grop_hc(rsum,rw,4,op,edge);
835827bd09bSSatish Balay 	  rsum[0]+=0.1; rsum[1]+=0.1; rsum[2]+=0.1; rsum[3]+=0.1;
836827bd09bSSatish Balay 
837827bd09bSSatish Balay 	  if (id<mask)
838827bd09bSSatish Balay 	    {
839827bd09bSSatish Balay 	      /* mark dofs I own that have signal and not in sep set */
840827bd09bSSatish Balay 	      for (ct=i=0;i<n;i++)
841827bd09bSSatish Balay 		{
842827bd09bSSatish Balay 		  if ((!used[i])&&(lhs[i]!=0.0))
843827bd09bSSatish Balay 		    {
844827bd09bSSatish Balay 		      ct++; nfo++;
845827bd09bSSatish Balay 
846827bd09bSSatish Balay 		      if (nfo>n)
847827bd09bSSatish Balay 			{error_msg_fatal("nfo about to exceed n\n");}
848827bd09bSSatish Balay 
849827bd09bSSatish Balay 		      *--iptr = local2global[i];
850827bd09bSSatish Balay 		      used[i]=edge;
851827bd09bSSatish Balay 		    }
852827bd09bSSatish Balay 		}
853827bd09bSSatish Balay 	      if (ct>1) {ivec_sort(iptr,ct);}
854827bd09bSSatish Balay 
855827bd09bSSatish Balay 	      lnsep[edge]=ct;
856827bd09bSSatish Balay 	      nsep[edge]=(int) rsum[0];
857827bd09bSSatish Balay 	      dir [edge]=LEFT;
858827bd09bSSatish Balay 	    }
859827bd09bSSatish Balay 
860827bd09bSSatish Balay 	  if (id>=mask)
861827bd09bSSatish Balay 	    {
862827bd09bSSatish Balay 	      /* mark dofs I own that have signal and not in sep set */
863827bd09bSSatish Balay 	      for (ct=i=0;i<n;i++)
864827bd09bSSatish Balay 		{
865827bd09bSSatish Balay 		  if ((!used[i])&&(rhs[i]!=0.0))
866827bd09bSSatish Balay 		    {
867827bd09bSSatish Balay 		      ct++; nfo++;
868827bd09bSSatish Balay 
869827bd09bSSatish Balay 		      if (nfo>n)
870827bd09bSSatish Balay 			{error_msg_fatal("nfo about to exceed n\n");}
871827bd09bSSatish Balay 
872827bd09bSSatish Balay 		      *--iptr = local2global[i];
873827bd09bSSatish Balay 		      used[i]=edge;
874827bd09bSSatish Balay 		    }
875827bd09bSSatish Balay 		}
876827bd09bSSatish Balay 	      if (ct>1) {ivec_sort(iptr,ct);}
877827bd09bSSatish Balay 
878827bd09bSSatish Balay 	      lnsep[edge]=ct;
879827bd09bSSatish Balay 	      nsep[edge]= (int) rsum[1];
880827bd09bSSatish Balay 	      dir [edge]=RIGHT;
881827bd09bSSatish Balay 	    }
882827bd09bSSatish Balay 
883827bd09bSSatish Balay 	  /* LATER or we can recur on these to order seps at this level */
884827bd09bSSatish Balay 	  /* do we need full set of separators for this?                */
885827bd09bSSatish Balay 
886827bd09bSSatish Balay 	  /* fold rhs hc into lower */
887827bd09bSSatish Balay 	  if (id>=mask)
888827bd09bSSatish Balay 	    {id-=mask;}
889827bd09bSSatish Balay 	}
890827bd09bSSatish Balay     }
891827bd09bSSatish Balay   else
892827bd09bSSatish Balay     {
893827bd09bSSatish Balay       for (iptr=fo+n,id=my_id,mask=num_nodes>>1,edge=level;edge>0;edge--,mask>>=1)
894827bd09bSSatish Balay 	{
895827bd09bSSatish Balay 	  /* set rsh of hc, fire, and collect lhs responses */
896827bd09bSSatish Balay 	  (id<mask) ? rvec_zero(lhs,m) : rvec_set(lhs,1.0,m);
897827bd09bSSatish Balay 	  gs_gop_hc(gs_handle,lhs,"+\0",edge);
898827bd09bSSatish Balay 
899827bd09bSSatish Balay 	  /* set lsh of hc, fire, and collect rhs responses */
900827bd09bSSatish Balay 	  (id<mask) ? rvec_set(rhs,1.0,m) : rvec_zero(rhs,m);
901827bd09bSSatish Balay 	  gs_gop_hc(gs_handle,rhs,"+\0",edge);
902827bd09bSSatish Balay 
903827bd09bSSatish Balay 	  /* count number of dofs I own that have signal and not in sep set */
904827bd09bSSatish Balay 	  for (ivec_zero(sum,4),ct=i=0;i<n;i++)
905827bd09bSSatish Balay 	    {
906827bd09bSSatish Balay 	      if (!used[i])
907827bd09bSSatish Balay 		{
908827bd09bSSatish Balay 		  /* number of unmarked dofs on node */
909827bd09bSSatish Balay 		  ct++;
910827bd09bSSatish Balay 		  /* number of dofs to be marked on lhs hc */
911827bd09bSSatish Balay 		  if ((id< mask)&&(lhs[i]!=0.0)) {sum[0]++;}
912827bd09bSSatish Balay 		  /* number of dofs to be marked on rhs hc */
913827bd09bSSatish Balay 		  if ((id>=mask)&&(rhs[i]!=0.0)) {sum[1]++;}
914827bd09bSSatish Balay 		}
915827bd09bSSatish Balay 	    }
916827bd09bSSatish Balay 
917827bd09bSSatish Balay 	  /* go for load balance - choose half with most unmarked dofs, bias LHS */
918827bd09bSSatish Balay 	  (id<mask) ? (sum[2]=ct) : (sum[3]=ct);
919827bd09bSSatish Balay 	  giop_hc(sum,w,4,op,edge);
920827bd09bSSatish Balay 
921827bd09bSSatish Balay 	  /* lhs hc wins */
922827bd09bSSatish Balay 	  if (sum[2]>=sum[3])
923827bd09bSSatish Balay 	    {
924827bd09bSSatish Balay 	      if (id<mask)
925827bd09bSSatish Balay 		{
926827bd09bSSatish Balay 		  /* mark dofs I own that have signal and not in sep set */
927827bd09bSSatish Balay 		  for (ct=i=0;i<n;i++)
928827bd09bSSatish Balay 		    {
929827bd09bSSatish Balay 		      if ((!used[i])&&(lhs[i]!=0.0))
930827bd09bSSatish Balay 			{
931827bd09bSSatish Balay 			  ct++; nfo++;
932827bd09bSSatish Balay 			  *--iptr = local2global[i];
933827bd09bSSatish Balay 			  used[i]=edge;
934827bd09bSSatish Balay 			}
935827bd09bSSatish Balay 		    }
936827bd09bSSatish Balay 		  if (ct>1) {ivec_sort(iptr,ct);}
937827bd09bSSatish Balay 		  lnsep[edge]=ct;
938827bd09bSSatish Balay 		}
939827bd09bSSatish Balay 	      nsep[edge]=sum[0];
940827bd09bSSatish Balay 	      dir [edge]=LEFT;
941827bd09bSSatish Balay 	    }
942827bd09bSSatish Balay 	  /* rhs hc wins */
943827bd09bSSatish Balay 	  else
944827bd09bSSatish Balay 	    {
945827bd09bSSatish Balay 	      if (id>=mask)
946827bd09bSSatish Balay 		{
947827bd09bSSatish Balay 		  /* mark dofs I own that have signal and not in sep set */
948827bd09bSSatish Balay 		  for (ct=i=0;i<n;i++)
949827bd09bSSatish Balay 		    {
950827bd09bSSatish Balay 		      if ((!used[i])&&(rhs[i]!=0.0))
951827bd09bSSatish Balay 			{
952827bd09bSSatish Balay 			  ct++; nfo++;
953827bd09bSSatish Balay 			  *--iptr = local2global[i];
954827bd09bSSatish Balay 			  used[i]=edge;
955827bd09bSSatish Balay 			}
956827bd09bSSatish Balay 		    }
957827bd09bSSatish Balay 		  if (ct>1) {ivec_sort(iptr,ct);}
958827bd09bSSatish Balay 		  lnsep[edge]=ct;
959827bd09bSSatish Balay 		}
960827bd09bSSatish Balay 	      nsep[edge]=sum[1];
961827bd09bSSatish Balay 	      dir [edge]=RIGHT;
962827bd09bSSatish Balay 	    }
963827bd09bSSatish Balay 	  /* LATER or we can recur on these to order seps at this level */
964827bd09bSSatish Balay 	  /* do we need full set of separators for this?                */
965827bd09bSSatish Balay 
966827bd09bSSatish Balay 	  /* fold rhs hc into lower */
967827bd09bSSatish Balay 	  if (id>=mask)
968827bd09bSSatish Balay 	    {id-=mask;}
969827bd09bSSatish Balay 	}
970827bd09bSSatish Balay     }
971827bd09bSSatish Balay 
972827bd09bSSatish Balay 
973827bd09bSSatish Balay   /* level 0 is on processor case - so mark the remainder */
974827bd09bSSatish Balay   for (ct=i=0;i<n;i++)
975827bd09bSSatish Balay     {
976827bd09bSSatish Balay       if (!used[i])
977827bd09bSSatish Balay 	{
978827bd09bSSatish Balay 	  ct++; nfo++;
979827bd09bSSatish Balay 	  *--iptr = local2global[i];
980827bd09bSSatish Balay 	  used[i]=edge;
981827bd09bSSatish Balay 	}
982827bd09bSSatish Balay     }
983827bd09bSSatish Balay   if (ct>1) {ivec_sort(iptr,ct);}
984827bd09bSSatish Balay   lnsep[edge]=ct;
985827bd09bSSatish Balay   nsep [edge]=ct;
986827bd09bSSatish Balay   dir  [edge]=LEFT;
987827bd09bSSatish Balay 
988827bd09bSSatish Balay   xxt_handle->info->nsep=nsep;
989827bd09bSSatish Balay   xxt_handle->info->lnsep=lnsep;
990827bd09bSSatish Balay   xxt_handle->info->fo=fo;
991827bd09bSSatish Balay   xxt_handle->info->nfo=nfo;
992827bd09bSSatish Balay 
993*a501084fSBarry Smith   free(dir);
994*a501084fSBarry Smith   free(lhs);
995*a501084fSBarry Smith   free(rhs);
996*a501084fSBarry Smith   free(used);
997827bd09bSSatish Balay 
998827bd09bSSatish Balay }
999827bd09bSSatish Balay 
1000827bd09bSSatish Balay 
1001827bd09bSSatish Balay /*************************************xxt.c************************************
1002827bd09bSSatish Balay Function: set_mvi
1003827bd09bSSatish Balay 
1004827bd09bSSatish Balay Input :
1005827bd09bSSatish Balay Output:
1006827bd09bSSatish Balay Return:
1007827bd09bSSatish Balay Description:
1008827bd09bSSatish Balay **************************************xxt.c***********************************/
1009827bd09bSSatish Balay static
1010827bd09bSSatish Balay mv_info *set_mvi(int *local2global, int n, int m, void *matvec, void *grid_data)
1011827bd09bSSatish Balay {
1012827bd09bSSatish Balay   mv_info *mvi;
1013827bd09bSSatish Balay 
1014827bd09bSSatish Balay 
1015*a501084fSBarry Smith   mvi = (mv_info*)malloc(sizeof(mv_info));
1016827bd09bSSatish Balay   mvi->n=n;
1017827bd09bSSatish Balay   mvi->m=m;
1018827bd09bSSatish Balay   mvi->n_global=-1;
1019827bd09bSSatish Balay   mvi->m_global=-1;
1020*a501084fSBarry Smith   mvi->local2global=(int*)malloc((m+1)*sizeof(PetscInt));
1021827bd09bSSatish Balay   ivec_copy(mvi->local2global,local2global,m);
1022827bd09bSSatish Balay   mvi->local2global[m] = INT_MAX;
1023*a501084fSBarry Smith   mvi->matvec=(PetscErrorCode (*)(mv_info*,PetscScalar*,PetscScalar*))matvec;
1024827bd09bSSatish Balay   mvi->grid_data=grid_data;
1025827bd09bSSatish Balay 
1026827bd09bSSatish Balay   /* set xxt communication handle to perform restricted matvec */
1027827bd09bSSatish Balay   mvi->gs_handle = gs_init(local2global, m, num_nodes);
1028827bd09bSSatish Balay 
1029827bd09bSSatish Balay   return(mvi);
1030827bd09bSSatish Balay }
1031827bd09bSSatish Balay 
1032827bd09bSSatish Balay 
1033827bd09bSSatish Balay /*************************************xxt.c************************************
1034827bd09bSSatish Balay Function: set_mvi
1035827bd09bSSatish Balay 
1036827bd09bSSatish Balay Input :
1037827bd09bSSatish Balay Output:
1038827bd09bSSatish Balay Return:
1039827bd09bSSatish Balay Description:
1040827bd09bSSatish Balay 
1041827bd09bSSatish Balay       computes u = A.v
1042827bd09bSSatish Balay       do_matvec(xxt_handle->mvi,v,u);
1043827bd09bSSatish Balay **************************************xxt.c***********************************/
1044827bd09bSSatish Balay static
1045*a501084fSBarry Smith void do_matvec(mv_info *A, PetscScalar *v, PetscScalar *u)
1046827bd09bSSatish Balay {
1047827bd09bSSatish Balay   A->matvec((mv_info*)A->grid_data,v,u);
1048827bd09bSSatish Balay }
1049827bd09bSSatish Balay 
1050827bd09bSSatish Balay 
1051827bd09bSSatish Balay 
1052