xref: /petsc/src/ksp/pc/impls/tfs/gs.c (revision 6b9672284b0484dde10212faecc0be7fd83e3315)
1827bd09bSSatish Balay 
2827bd09bSSatish Balay /***********************************gs.c***************************************
3827bd09bSSatish Balay 
4827bd09bSSatish Balay Author: Henry M. Tufo III
5827bd09bSSatish Balay 
6827bd09bSSatish Balay e-mail: hmt@cs.brown.edu
7827bd09bSSatish Balay 
8827bd09bSSatish Balay snail-mail:
9827bd09bSSatish Balay Division of Applied Mathematics
10827bd09bSSatish Balay Brown University
11827bd09bSSatish Balay Providence, RI 02912
12827bd09bSSatish Balay 
13827bd09bSSatish Balay Last Modification:
14827bd09bSSatish Balay 6.21.97
15827bd09bSSatish Balay ************************************gs.c**************************************/
16827bd09bSSatish Balay 
17827bd09bSSatish Balay /***********************************gs.c***************************************
18827bd09bSSatish Balay File Description:
19827bd09bSSatish Balay -----------------
20827bd09bSSatish Balay 
21827bd09bSSatish Balay ************************************gs.c**************************************/
22827bd09bSSatish Balay 
23c6db04a5SJed Brown #include <../src/ksp/pc/impls/tfs/tfs.h>
2439945688SSatish Balay 
25827bd09bSSatish Balay /* default length of number of items via tree - doubles if exceeded */
26827bd09bSSatish Balay #define TREE_BUF_SZ 2048;
27827bd09bSSatish Balay #define GS_VEC_SZ   1
28827bd09bSSatish Balay 
29827bd09bSSatish Balay 
30827bd09bSSatish Balay 
31827bd09bSSatish Balay /***********************************gs.c***************************************
32827bd09bSSatish Balay Type: struct gather_scatter_id
33827bd09bSSatish Balay ------------------------------
34827bd09bSSatish Balay 
35827bd09bSSatish Balay ************************************gs.c**************************************/
36827bd09bSSatish Balay typedef struct gather_scatter_id {
3752f87cdaSBarry Smith   PetscInt    id;
3852f87cdaSBarry Smith   PetscInt    nel_min;
3952f87cdaSBarry Smith   PetscInt    nel_max;
4052f87cdaSBarry Smith   PetscInt    nel_sum;
4152f87cdaSBarry Smith   PetscInt    negl;
4252f87cdaSBarry Smith   PetscInt    gl_max;
4352f87cdaSBarry Smith   PetscInt    gl_min;
4452f87cdaSBarry Smith   PetscInt    repeats;
4552f87cdaSBarry Smith   PetscInt    ordered;
4652f87cdaSBarry Smith   PetscInt    positive;
47a501084fSBarry Smith   PetscScalar *vals;
48827bd09bSSatish Balay 
49827bd09bSSatish Balay   /* bit mask info */
5052f87cdaSBarry Smith   PetscInt *my_proc_mask;
5152f87cdaSBarry Smith   PetscInt mask_sz;
5252f87cdaSBarry Smith   PetscInt *ngh_buf;
5352f87cdaSBarry Smith   PetscInt ngh_buf_sz;
5452f87cdaSBarry Smith   PetscInt *nghs;
5552f87cdaSBarry Smith   PetscInt num_nghs;
5652f87cdaSBarry Smith   PetscInt max_nghs;
5752f87cdaSBarry Smith   PetscInt *pw_nghs;
5852f87cdaSBarry Smith   PetscInt num_pw_nghs;
5952f87cdaSBarry Smith   PetscInt *tree_nghs;
6052f87cdaSBarry Smith   PetscInt num_tree_nghs;
61827bd09bSSatish Balay 
6252f87cdaSBarry Smith   PetscInt num_loads;
63827bd09bSSatish Balay 
64827bd09bSSatish Balay   /* repeats == true -> local info */
6552f87cdaSBarry Smith   PetscInt nel;         /* number of unique elememts */
6652f87cdaSBarry Smith   PetscInt *elms;       /* of size nel */
6752f87cdaSBarry Smith   PetscInt nel_total;
6852f87cdaSBarry Smith   PetscInt *local_elms; /* of size nel_total */
6952f87cdaSBarry Smith   PetscInt *companion;  /* of size nel_total */
70827bd09bSSatish Balay 
71827bd09bSSatish Balay   /* local info */
7252f87cdaSBarry Smith   PetscInt num_local_total;
7352f87cdaSBarry Smith   PetscInt local_strength;
7452f87cdaSBarry Smith   PetscInt num_local;
7552f87cdaSBarry Smith   PetscInt *num_local_reduce;
7652f87cdaSBarry Smith   PetscInt **local_reduce;
7752f87cdaSBarry Smith   PetscInt num_local_gop;
7852f87cdaSBarry Smith   PetscInt *num_gop_local_reduce;
7952f87cdaSBarry Smith   PetscInt **gop_local_reduce;
80827bd09bSSatish Balay 
81827bd09bSSatish Balay   /* pairwise info */
8252f87cdaSBarry Smith   PetscInt    level;
8352f87cdaSBarry Smith   PetscInt    num_pairs;
8452f87cdaSBarry Smith   PetscInt    max_pairs;
8552f87cdaSBarry Smith   PetscInt    loc_node_pairs;
8652f87cdaSBarry Smith   PetscInt    max_node_pairs;
8752f87cdaSBarry Smith   PetscInt    min_node_pairs;
8852f87cdaSBarry Smith   PetscInt    avg_node_pairs;
8952f87cdaSBarry Smith   PetscInt    *pair_list;
9052f87cdaSBarry Smith   PetscInt    *msg_sizes;
9152f87cdaSBarry Smith   PetscInt    **node_list;
9252f87cdaSBarry Smith   PetscInt    len_pw_list;
9352f87cdaSBarry Smith   PetscInt    *pw_elm_list;
94a501084fSBarry Smith   PetscScalar *pw_vals;
95827bd09bSSatish Balay 
96827bd09bSSatish Balay   MPI_Request *msg_ids_in;
97827bd09bSSatish Balay   MPI_Request *msg_ids_out;
98827bd09bSSatish Balay 
99a501084fSBarry Smith   PetscScalar *out;
100a501084fSBarry Smith   PetscScalar *in;
10152f87cdaSBarry Smith   PetscInt    msg_total;
102827bd09bSSatish Balay 
103827bd09bSSatish Balay   /* tree - crystal accumulator info */
10452f87cdaSBarry Smith   PetscInt max_left_over;
10552f87cdaSBarry Smith   PetscInt *pre;
10652f87cdaSBarry Smith   PetscInt *in_num;
10752f87cdaSBarry Smith   PetscInt *out_num;
10852f87cdaSBarry Smith   PetscInt **in_list;
10952f87cdaSBarry Smith   PetscInt **out_list;
110827bd09bSSatish Balay 
111827bd09bSSatish Balay   /* new tree work*/
11252f87cdaSBarry Smith   PetscInt    tree_nel;
11352f87cdaSBarry Smith   PetscInt    *tree_elms;
114a501084fSBarry Smith   PetscScalar *tree_buf;
115a501084fSBarry Smith   PetscScalar *tree_work;
116827bd09bSSatish Balay 
11752f87cdaSBarry Smith   PetscInt tree_map_sz;
11852f87cdaSBarry Smith   PetscInt *tree_map_in;
11952f87cdaSBarry Smith   PetscInt *tree_map_out;
120827bd09bSSatish Balay 
121827bd09bSSatish Balay   /* current memory status */
12252f87cdaSBarry Smith   PetscInt gl_bss_min;
12352f87cdaSBarry Smith   PetscInt gl_perm_min;
124827bd09bSSatish Balay 
125ca8e9878SJed Brown   /* max segment size for PCTFS_gs_gop_vec() */
12652f87cdaSBarry Smith   PetscInt vec_sz;
127827bd09bSSatish Balay 
128827bd09bSSatish Balay   /* hack to make paul happy */
129ca8e9878SJed Brown   MPI_Comm PCTFS_gs_comm;
130827bd09bSSatish Balay 
131ca8e9878SJed Brown } PCTFS_gs_id;
132827bd09bSSatish Balay 
133ca8e9878SJed Brown static PCTFS_gs_id *gsi_check_args(PetscInt *elms, PetscInt nel, PetscInt level);
134ca8e9878SJed Brown static PetscErrorCode gsi_via_bit_mask(PCTFS_gs_id *gs);
135ca8e9878SJed Brown static PetscErrorCode get_ngh_buf(PCTFS_gs_id *gs);
136ca8e9878SJed Brown static PetscErrorCode set_pairwise(PCTFS_gs_id *gs);
137ca8e9878SJed Brown static PCTFS_gs_id *gsi_new(void);
138ca8e9878SJed Brown static PetscErrorCode set_tree(PCTFS_gs_id *gs);
139827bd09bSSatish Balay 
140827bd09bSSatish Balay /* same for all but vector flavor */
141ca8e9878SJed Brown static PetscErrorCode PCTFS_gs_gop_local_out(PCTFS_gs_id *gs, PetscScalar *vals);
142827bd09bSSatish Balay /* vector flavor */
143ca8e9878SJed Brown static PetscErrorCode PCTFS_gs_gop_vec_local_out(PCTFS_gs_id *gs, PetscScalar *vals, PetscInt step);
144827bd09bSSatish Balay 
145ca8e9878SJed Brown static PetscErrorCode PCTFS_gs_gop_vec_plus(PCTFS_gs_id *gs, PetscScalar *in_vals, PetscInt step);
146ca8e9878SJed Brown static PetscErrorCode PCTFS_gs_gop_vec_pairwise_plus(PCTFS_gs_id *gs, PetscScalar *in_vals, PetscInt step);
147ca8e9878SJed Brown static PetscErrorCode PCTFS_gs_gop_vec_local_plus(PCTFS_gs_id *gs, PetscScalar *vals, PetscInt step);
148ca8e9878SJed Brown static PetscErrorCode PCTFS_gs_gop_vec_local_in_plus(PCTFS_gs_id *gs, PetscScalar *vals, PetscInt step);
149ca8e9878SJed Brown static PetscErrorCode PCTFS_gs_gop_vec_tree_plus(PCTFS_gs_id *gs, PetscScalar *vals, PetscInt step);
150827bd09bSSatish Balay 
151827bd09bSSatish Balay 
152ca8e9878SJed Brown static PetscErrorCode PCTFS_gs_gop_local_plus(PCTFS_gs_id *gs, PetscScalar *vals);
153ca8e9878SJed Brown static PetscErrorCode PCTFS_gs_gop_local_in_plus(PCTFS_gs_id *gs, PetscScalar *vals);
154827bd09bSSatish Balay 
155ca8e9878SJed Brown static PetscErrorCode PCTFS_gs_gop_plus_hc(PCTFS_gs_id *gs, PetscScalar *in_vals, PetscInt dim);
156ca8e9878SJed Brown static PetscErrorCode PCTFS_gs_gop_pairwise_plus_hc(PCTFS_gs_id *gs, PetscScalar *in_vals, PetscInt dim);
157ca8e9878SJed Brown static PetscErrorCode PCTFS_gs_gop_tree_plus_hc(PCTFS_gs_id *gs, PetscScalar *vals, PetscInt dim);
158827bd09bSSatish Balay 
159827bd09bSSatish Balay /* global vars */
160827bd09bSSatish Balay /* from comm.c module */
161827bd09bSSatish Balay 
16252f87cdaSBarry Smith static PetscInt num_gs_ids = 0;
163827bd09bSSatish Balay 
164827bd09bSSatish Balay /* should make this dynamic ... later */
16552f87cdaSBarry Smith static PetscInt msg_buf    =MAX_MSG_BUF;
16652f87cdaSBarry Smith static PetscInt vec_sz     =GS_VEC_SZ;
16752f87cdaSBarry Smith static PetscInt *tree_buf  =NULL;
16852f87cdaSBarry Smith static PetscInt tree_buf_sz=0;
16952f87cdaSBarry Smith static PetscInt ntree      =0;
170827bd09bSSatish Balay 
171f1ed62a8SBarry Smith /***************************************************************************/
172ca8e9878SJed Brown PetscErrorCode PCTFS_gs_init_vec_sz(PetscInt size)
173827bd09bSSatish Balay {
1743fdc5746SBarry Smith   PetscFunctionBegin;
175827bd09bSSatish Balay   vec_sz = size;
1763fdc5746SBarry Smith   PetscFunctionReturn(0);
177827bd09bSSatish Balay }
178827bd09bSSatish Balay 
179f1ed62a8SBarry Smith /******************************************************************************/
180ca8e9878SJed Brown PetscErrorCode PCTFS_gs_init_msg_buf_sz(PetscInt buf_size)
181827bd09bSSatish Balay {
1823fdc5746SBarry Smith   PetscFunctionBegin;
183827bd09bSSatish Balay   msg_buf = buf_size;
1843fdc5746SBarry Smith   PetscFunctionReturn(0);
185827bd09bSSatish Balay }
186827bd09bSSatish Balay 
187f1ed62a8SBarry Smith /******************************************************************************/
188ca8e9878SJed Brown PCTFS_gs_id *PCTFS_gs_init(PetscInt *elms, PetscInt nel, PetscInt level)
189827bd09bSSatish Balay {
190ca8e9878SJed Brown   PCTFS_gs_id    *gs;
191ca8e9878SJed Brown   MPI_Group      PCTFS_gs_group;
192ca8e9878SJed Brown   MPI_Comm       PCTFS_gs_comm;
193f1ed62a8SBarry Smith   PetscErrorCode ierr;
194827bd09bSSatish Balay 
1953fdc5746SBarry Smith   PetscFunctionBegin;
196827bd09bSSatish Balay   /* ensure that communication package has been initialized */
197b1c944f5SJed Brown   PCTFS_comm_init();
198827bd09bSSatish Balay 
199827bd09bSSatish Balay 
200827bd09bSSatish Balay   /* determines if we have enough dynamic/semi-static memory */
201827bd09bSSatish Balay   /* checks input, allocs and sets gd_id template            */
202827bd09bSSatish Balay   gs = gsi_check_args(elms,nel,level);
203827bd09bSSatish Balay 
204827bd09bSSatish Balay   /* only bit mask version up and working for the moment    */
205827bd09bSSatish Balay   /* LATER :: get int list version working for sparse pblms */
206f1ed62a8SBarry Smith   ierr = gsi_via_bit_mask(gs);CHKERRABORT(PETSC_COMM_WORLD,ierr);
207827bd09bSSatish Balay 
208827bd09bSSatish Balay 
209ca8e9878SJed Brown   ierr = MPI_Comm_group(MPI_COMM_WORLD,&PCTFS_gs_group);CHKERRABORT(PETSC_COMM_WORLD,ierr);
210ca8e9878SJed Brown   ierr = MPI_Comm_create(MPI_COMM_WORLD,PCTFS_gs_group,&PCTFS_gs_comm);CHKERRABORT(PETSC_COMM_WORLD,ierr);
211*6b967228SBarry Smith   ierr = MPI_Group_free(&PCTFS_gs_group);CHKERRABORT(PETSC_COMM_WORLD,ierr);
2122fa5cd67SKarl Rupp 
213ca8e9878SJed Brown   gs->PCTFS_gs_comm=PCTFS_gs_comm;
214827bd09bSSatish Balay 
215827bd09bSSatish Balay   return(gs);
216827bd09bSSatish Balay }
217827bd09bSSatish Balay 
218f1ed62a8SBarry Smith /******************************************************************************/
219ca8e9878SJed Brown static PCTFS_gs_id *gsi_new(void)
220827bd09bSSatish Balay {
221f1ed62a8SBarry Smith   PetscErrorCode ierr;
222ca8e9878SJed Brown   PCTFS_gs_id    *gs;
223ca8e9878SJed Brown   gs   = (PCTFS_gs_id*) malloc(sizeof(PCTFS_gs_id));
224ca8e9878SJed Brown   ierr = PetscMemzero(gs,sizeof(PCTFS_gs_id));CHKERRABORT(PETSC_COMM_WORLD,ierr);
225827bd09bSSatish Balay   return(gs);
226827bd09bSSatish Balay }
227827bd09bSSatish Balay 
228f1ed62a8SBarry Smith /******************************************************************************/
229ca8e9878SJed Brown static PCTFS_gs_id *gsi_check_args(PetscInt *in_elms, PetscInt nel, PetscInt level)
230827bd09bSSatish Balay {
23152f87cdaSBarry Smith   PetscInt       i, j, k, t2;
23252f87cdaSBarry Smith   PetscInt       *companion, *elms, *unique, *iptr;
23352f87cdaSBarry Smith   PetscInt       num_local=0, *num_to_reduce, **local_reduce;
23452f87cdaSBarry Smith   PetscInt       oprs[]   = {NON_UNIFORM,GL_MIN,GL_MAX,GL_ADD,GL_MIN,GL_MAX,GL_MIN,GL_B_AND};
23552f87cdaSBarry Smith   PetscInt       vals[sizeof(oprs)/sizeof(oprs[0])-1];
23652f87cdaSBarry Smith   PetscInt       work[sizeof(oprs)/sizeof(oprs[0])-1];
237ca8e9878SJed Brown   PCTFS_gs_id    *gs;
238d1528f56SBarry Smith   PetscErrorCode ierr;
239827bd09bSSatish Balay 
240827bd09bSSatish Balay 
241c1235816SBarry Smith   if (!in_elms) SETERRABORT(PETSC_COMM_WORLD,PETSC_ERR_PLIB,"elms point to nothing!!!\n");
242c1235816SBarry Smith   if (nel<0)    SETERRABORT(PETSC_COMM_WORLD,PETSC_ERR_PLIB,"can't have fewer than 0 elms!!!\n");
243827bd09bSSatish Balay 
244db4deed7SKarl Rupp   if (nel==0) { ierr = PetscInfo(0,"I don't have any elements!!!\n");CHKERRABORT(PETSC_COMM_WORLD,ierr); }
245827bd09bSSatish Balay 
246827bd09bSSatish Balay   /* get space for gs template */
247827bd09bSSatish Balay   gs     = gsi_new();
248827bd09bSSatish Balay   gs->id = ++num_gs_ids;
249827bd09bSSatish Balay 
250827bd09bSSatish Balay   /* hmt 6.4.99                                            */
251827bd09bSSatish Balay   /* caller can set global ids that don't participate to 0 */
252ca8e9878SJed Brown   /* PCTFS_gs_init ignores all zeros in elm list                 */
253827bd09bSSatish Balay   /* negative global ids are still invalid                 */
2542fa5cd67SKarl Rupp   for (i=j=0; i<nel; i++) {
2552fa5cd67SKarl Rupp     if (in_elms[i]!=0) j++;
2562fa5cd67SKarl Rupp   }
257827bd09bSSatish Balay 
258827bd09bSSatish Balay   k=nel; nel=j;
259827bd09bSSatish Balay 
260827bd09bSSatish Balay   /* copy over in_elms list and create inverse map */
26152f87cdaSBarry Smith   elms      = (PetscInt*) malloc((nel+1)*sizeof(PetscInt));
26252f87cdaSBarry Smith   companion = (PetscInt*) malloc(nel*sizeof(PetscInt));
2631d7d0905SBarry Smith 
264db4deed7SKarl Rupp   for (i=j=0; i<k; i++) {
265db4deed7SKarl Rupp     if (in_elms[i]!=0) { elms[j] = in_elms[i]; companion[j++] = i; }
266827bd09bSSatish Balay   }
267827bd09bSSatish Balay 
268c1235816SBarry Smith   if (j!=nel) SETERRABORT(PETSC_COMM_WORLD,PETSC_ERR_PLIB,"nel j mismatch!\n");
269827bd09bSSatish Balay 
270827bd09bSSatish Balay   /* pre-pass ... check to see if sorted */
271827bd09bSSatish Balay   elms[nel] = INT_MAX;
272827bd09bSSatish Balay   iptr      = elms;
273827bd09bSSatish Balay   unique    = elms+1;
274827bd09bSSatish Balay   j         =0;
275db4deed7SKarl Rupp   while (*iptr!=INT_MAX) {
276db4deed7SKarl Rupp     if (*iptr++>*unique++) { j=1; break; }
277827bd09bSSatish Balay   }
278827bd09bSSatish Balay 
279827bd09bSSatish Balay   /* set up inverse map */
280db4deed7SKarl Rupp   if (j) {
281f1ed62a8SBarry Smith     ierr = PetscInfo(0,"gsi_check_args() :: elm list *not* sorted!\n");CHKERRABORT(PETSC_COMM_WORLD,ierr);
282ca8e9878SJed Brown     ierr = PCTFS_SMI_sort((void*)elms, (void*)companion, nel, SORT_INTEGER);CHKERRABORT(PETSC_COMM_WORLD,ierr);
2832fa5cd67SKarl Rupp   } else { ierr = PetscInfo(0,"gsi_check_args() :: elm list sorted!\n");CHKERRABORT(PETSC_COMM_WORLD,ierr); }
284827bd09bSSatish Balay   elms[nel] = INT_MIN;
285827bd09bSSatish Balay 
286827bd09bSSatish Balay   /* first pass */
287827bd09bSSatish Balay   /* determine number of unique elements, check pd */
288db4deed7SKarl Rupp   for (i=k=0; i<nel; i+=j) {
289827bd09bSSatish Balay     t2 = elms[i];
290827bd09bSSatish Balay     j  = ++i;
291827bd09bSSatish Balay 
292827bd09bSSatish Balay     /* clump 'em for now */
2932fa5cd67SKarl Rupp     while (elms[j]==t2) j++;
294827bd09bSSatish Balay 
295827bd09bSSatish Balay     /* how many together and num local */
296db4deed7SKarl Rupp     if (j-=i) { num_local++; k+=j; }
297827bd09bSSatish Balay   }
298827bd09bSSatish Balay 
299827bd09bSSatish Balay   /* how many unique elements? */
300827bd09bSSatish Balay   gs->repeats = k;
301827bd09bSSatish Balay   gs->nel     = nel-k;
302827bd09bSSatish Balay 
303827bd09bSSatish Balay 
304827bd09bSSatish Balay   /* number of repeats? */
305827bd09bSSatish Balay   gs->num_local        = num_local;
306827bd09bSSatish Balay   num_local           += 2;
30752f87cdaSBarry Smith   gs->local_reduce     = local_reduce=(PetscInt**)malloc(num_local*sizeof(PetscInt*));
30852f87cdaSBarry Smith   gs->num_local_reduce = num_to_reduce=(PetscInt*) malloc(num_local*sizeof(PetscInt));
309827bd09bSSatish Balay 
31052f87cdaSBarry Smith   unique         = (PetscInt*) malloc((gs->nel+1)*sizeof(PetscInt));
311827bd09bSSatish Balay   gs->elms       = unique;
312827bd09bSSatish Balay   gs->nel_total  = nel;
313827bd09bSSatish Balay   gs->local_elms = elms;
314827bd09bSSatish Balay   gs->companion  = companion;
315827bd09bSSatish Balay 
316827bd09bSSatish Balay   /* compess map as well as keep track of local ops */
317db4deed7SKarl Rupp   for (num_local=i=j=0; i<gs->nel; i++) {
318827bd09bSSatish Balay     k            = j;
319827bd09bSSatish Balay     t2           = unique[i] = elms[j];
320827bd09bSSatish Balay     companion[i] = companion[j];
321827bd09bSSatish Balay 
3222fa5cd67SKarl Rupp     while (elms[j]==t2) j++;
323827bd09bSSatish Balay 
324db4deed7SKarl Rupp     if ((t2=(j-k))>1) {
325827bd09bSSatish Balay       /* number together */
326827bd09bSSatish Balay       num_to_reduce[num_local] = t2++;
3272fa5cd67SKarl Rupp 
32852f87cdaSBarry Smith       iptr = local_reduce[num_local++] = (PetscInt*)malloc(t2*sizeof(PetscInt));
329827bd09bSSatish Balay 
330827bd09bSSatish Balay       /* to use binary searching don't remap until we check intersection */
331827bd09bSSatish Balay       *iptr++ = i;
332827bd09bSSatish Balay 
333827bd09bSSatish Balay       /* note that we're skipping the first one */
3342fa5cd67SKarl Rupp       while (++k<j) *(iptr++) = companion[k];
335827bd09bSSatish Balay       *iptr = -1;
336827bd09bSSatish Balay     }
337827bd09bSSatish Balay   }
338827bd09bSSatish Balay 
339827bd09bSSatish Balay   /* sentinel for ngh_buf */
340827bd09bSSatish Balay   unique[gs->nel]=INT_MAX;
341827bd09bSSatish Balay 
342827bd09bSSatish Balay   /* for two partition sort hack */
343827bd09bSSatish Balay   num_to_reduce[num_local]   = 0;
344827bd09bSSatish Balay   local_reduce[num_local]    = NULL;
345827bd09bSSatish Balay   num_to_reduce[++num_local] = 0;
346827bd09bSSatish Balay   local_reduce[num_local]    = NULL;
347827bd09bSSatish Balay 
348827bd09bSSatish Balay   /* load 'em up */
349827bd09bSSatish Balay   /* note one extra to hold NON_UNIFORM flag!!! */
350827bd09bSSatish Balay   vals[2] = vals[1] = vals[0] = nel;
351db4deed7SKarl Rupp   if (gs->nel>0) {
3521d7d0905SBarry Smith     vals[3] = unique[0];
3531d7d0905SBarry Smith     vals[4] = unique[gs->nel-1];
354db4deed7SKarl Rupp   } else {
3551d7d0905SBarry Smith     vals[3] = INT_MAX;
3561d7d0905SBarry Smith     vals[4] = INT_MIN;
357827bd09bSSatish Balay   }
358827bd09bSSatish Balay   vals[5] = level;
359827bd09bSSatish Balay   vals[6] = num_gs_ids;
360827bd09bSSatish Balay 
361827bd09bSSatish Balay   /* GLOBAL: send 'em out */
362b1c944f5SJed Brown   ierr = PCTFS_giop(vals,work,sizeof(oprs)/sizeof(oprs[0])-1,oprs);CHKERRABORT(PETSC_COMM_WORLD,ierr);
363827bd09bSSatish Balay 
364827bd09bSSatish Balay   /* must be semi-pos def - only pairwise depends on this */
365827bd09bSSatish Balay   /* LATER - remove this restriction */
366c1235816SBarry Smith   if (vals[3]<0) SETERRABORT(PETSC_COMM_WORLD,PETSC_ERR_PLIB,"gsi_check_args() :: system not semi-pos def \n");
367c1235816SBarry Smith   if (vals[4]==INT_MAX) SETERRABORT(PETSC_COMM_WORLD,PETSC_ERR_PLIB,"gsi_check_args() :: system ub too large !\n");
368827bd09bSSatish Balay 
369827bd09bSSatish Balay   gs->nel_min = vals[0];
370827bd09bSSatish Balay   gs->nel_max = vals[1];
371827bd09bSSatish Balay   gs->nel_sum = vals[2];
372827bd09bSSatish Balay   gs->gl_min  = vals[3];
373827bd09bSSatish Balay   gs->gl_max  = vals[4];
374827bd09bSSatish Balay   gs->negl    = vals[4]-vals[3]+1;
375827bd09bSSatish Balay 
376c1235816SBarry Smith   if (gs->negl<=0) SETERRABORT(PETSC_COMM_WORLD,PETSC_ERR_PLIB,"gsi_check_args() :: system empty or neg :: %d\n");
377827bd09bSSatish Balay 
378827bd09bSSatish Balay   /* LATER :: add level == -1 -> program selects level */
3792fa5cd67SKarl Rupp   if (vals[5]<0) vals[5]=0;
3802fa5cd67SKarl Rupp   else if (vals[5]>PCTFS_num_nodes) vals[5]=PCTFS_num_nodes;
381827bd09bSSatish Balay   gs->level = vals[5];
382827bd09bSSatish Balay 
383827bd09bSSatish Balay   return(gs);
384827bd09bSSatish Balay }
385827bd09bSSatish Balay 
386f1ed62a8SBarry Smith /******************************************************************************/
387ca8e9878SJed Brown static PetscErrorCode gsi_via_bit_mask(PCTFS_gs_id *gs)
388827bd09bSSatish Balay {
38952f87cdaSBarry Smith   PetscInt       i, nel, *elms;
39052f87cdaSBarry Smith   PetscInt       t1;
39152f87cdaSBarry Smith   PetscInt       **reduce;
39252f87cdaSBarry Smith   PetscInt       *map;
393f1ed62a8SBarry Smith   PetscErrorCode ierr;
394827bd09bSSatish Balay 
395f1ed62a8SBarry Smith   PetscFunctionBegin;
396ca8e9878SJed Brown   /* totally local removes ... PCTFS_ct_bits == 0 */
397827bd09bSSatish Balay   get_ngh_buf(gs);
398827bd09bSSatish Balay 
39994dd86cdSBarry Smith   if (gs->level) set_pairwise(gs);
40094dd86cdSBarry Smith   if (gs->max_left_over) set_tree(gs);
401827bd09bSSatish Balay 
402827bd09bSSatish Balay   /* intersection local and pairwise/tree? */
403827bd09bSSatish Balay   gs->num_local_total      = gs->num_local;
404827bd09bSSatish Balay   gs->gop_local_reduce     = gs->local_reduce;
405827bd09bSSatish Balay   gs->num_gop_local_reduce = gs->num_local_reduce;
406827bd09bSSatish Balay 
407827bd09bSSatish Balay   map = gs->companion;
408827bd09bSSatish Balay 
409827bd09bSSatish Balay   /* is there any local compression */
410d890fc11SSatish Balay   if (!gs->num_local) {
411827bd09bSSatish Balay     gs->local_strength = NONE;
412827bd09bSSatish Balay     gs->num_local_gop  = 0;
413d890fc11SSatish Balay   } else {
414827bd09bSSatish Balay     /* ok find intersection */
415827bd09bSSatish Balay     map    = gs->companion;
416827bd09bSSatish Balay     reduce = gs->local_reduce;
4174a2f8832SBarry Smith     for (i=0, t1=0; i<gs->num_local; i++, reduce++) {
4184a2f8832SBarry Smith       if ((PCTFS_ivec_binary_search(**reduce,gs->pw_elm_list,gs->len_pw_list)>=0) || PCTFS_ivec_binary_search(**reduce,gs->tree_map_in,gs->tree_map_sz)>=0) {
419827bd09bSSatish Balay         t1++;
420e32f2f54SBarry Smith         if (gs->num_local_reduce[i]<=0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"nobody in list?");
421827bd09bSSatish Balay         gs->num_local_reduce[i] *= -1;
422827bd09bSSatish Balay       }
423827bd09bSSatish Balay       **reduce=map[**reduce];
424827bd09bSSatish Balay     }
425827bd09bSSatish Balay 
426827bd09bSSatish Balay     /* intersection is empty */
427db4deed7SKarl Rupp     if (!t1) {
428827bd09bSSatish Balay       gs->local_strength = FULL;
429827bd09bSSatish Balay       gs->num_local_gop  = 0;
430db4deed7SKarl Rupp     } else { /* intersection not empty */
431827bd09bSSatish Balay       gs->local_strength = PARTIAL;
4322fa5cd67SKarl Rupp 
433ca8e9878SJed Brown       ierr = PCTFS_SMI_sort((void*)gs->num_local_reduce, (void*)gs->local_reduce, gs->num_local + 1, SORT_INT_PTR);CHKERRQ(ierr);
434827bd09bSSatish Balay 
435827bd09bSSatish Balay       gs->num_local_gop        = t1;
436827bd09bSSatish Balay       gs->num_local_total      =  gs->num_local;
437827bd09bSSatish Balay       gs->num_local           -= t1;
438827bd09bSSatish Balay       gs->gop_local_reduce     = gs->local_reduce;
439827bd09bSSatish Balay       gs->num_gop_local_reduce = gs->num_local_reduce;
440827bd09bSSatish Balay 
4412fa5cd67SKarl Rupp       for (i=0; i<t1; i++) {
442e32f2f54SBarry Smith         if (gs->num_gop_local_reduce[i]>=0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"they aren't negative?");
443827bd09bSSatish Balay         gs->num_gop_local_reduce[i] *= -1;
444827bd09bSSatish Balay         gs->local_reduce++;
445827bd09bSSatish Balay         gs->num_local_reduce++;
446827bd09bSSatish Balay       }
447827bd09bSSatish Balay       gs->local_reduce++;
448827bd09bSSatish Balay       gs->num_local_reduce++;
449827bd09bSSatish Balay     }
450827bd09bSSatish Balay   }
451827bd09bSSatish Balay 
452827bd09bSSatish Balay   elms = gs->pw_elm_list;
453827bd09bSSatish Balay   nel  = gs->len_pw_list;
4542fa5cd67SKarl Rupp   for (i=0; i<nel; i++) elms[i] = map[elms[i]];
455827bd09bSSatish Balay 
456827bd09bSSatish Balay   elms = gs->tree_map_in;
457827bd09bSSatish Balay   nel  = gs->tree_map_sz;
4582fa5cd67SKarl Rupp   for (i=0; i<nel; i++) elms[i] = map[elms[i]];
459827bd09bSSatish Balay 
460827bd09bSSatish Balay   /* clean up */
461a501084fSBarry Smith   free((void*) gs->local_elms);
462a501084fSBarry Smith   free((void*) gs->companion);
463a501084fSBarry Smith   free((void*) gs->elms);
464a501084fSBarry Smith   free((void*) gs->ngh_buf);
465827bd09bSSatish Balay   gs->local_elms = gs->companion = gs->elms = gs->ngh_buf = NULL;
4663fdc5746SBarry Smith   PetscFunctionReturn(0);
467827bd09bSSatish Balay }
468827bd09bSSatish Balay 
469f1ed62a8SBarry Smith /******************************************************************************/
47052f87cdaSBarry Smith static PetscErrorCode place_in_tree(PetscInt elm)
471827bd09bSSatish Balay {
47252f87cdaSBarry Smith   PetscInt *tp, n;
473827bd09bSSatish Balay 
4743fdc5746SBarry Smith   PetscFunctionBegin;
4752fa5cd67SKarl Rupp   if (ntree==tree_buf_sz) {
476db4deed7SKarl Rupp     if (tree_buf_sz) {
477827bd09bSSatish Balay       tp           = tree_buf;
478827bd09bSSatish Balay       n            = tree_buf_sz;
479827bd09bSSatish Balay       tree_buf_sz<<=1;
48052f87cdaSBarry Smith       tree_buf     = (PetscInt*)malloc(tree_buf_sz*sizeof(PetscInt));
481ca8e9878SJed Brown       PCTFS_ivec_copy(tree_buf,tp,n);
482a501084fSBarry Smith       free(tp);
483db4deed7SKarl Rupp     } else {
484827bd09bSSatish Balay       tree_buf_sz = TREE_BUF_SZ;
48552f87cdaSBarry Smith       tree_buf    = (PetscInt*)malloc(tree_buf_sz*sizeof(PetscInt));
486827bd09bSSatish Balay     }
487827bd09bSSatish Balay   }
488827bd09bSSatish Balay 
489827bd09bSSatish Balay   tree_buf[ntree++] = elm;
4903fdc5746SBarry Smith   PetscFunctionReturn(0);
491827bd09bSSatish Balay }
492827bd09bSSatish Balay 
493f1ed62a8SBarry Smith /******************************************************************************/
494ca8e9878SJed Brown static PetscErrorCode get_ngh_buf(PCTFS_gs_id *gs)
495827bd09bSSatish Balay {
49652f87cdaSBarry Smith   PetscInt       i, j, npw=0, ntree_map=0;
49752f87cdaSBarry Smith   PetscInt       p_mask_size, ngh_buf_size, buf_size;
49852f87cdaSBarry Smith   PetscInt       *p_mask, *sh_proc_mask, *pw_sh_proc_mask;
49952f87cdaSBarry Smith   PetscInt       *ngh_buf, *buf1, *buf2;
50052f87cdaSBarry Smith   PetscInt       offset, per_load, num_loads, or_ct, start, end;
50152f87cdaSBarry Smith   PetscInt       *ptr1, *ptr2, i_start, negl, nel, *elms;
50252f87cdaSBarry Smith   PetscInt       oper=GL_B_OR;
50352f87cdaSBarry Smith   PetscInt       *ptr3, *t_mask, level, ct1, ct2;
504f1ed62a8SBarry Smith   PetscErrorCode ierr;
505827bd09bSSatish Balay 
5063fdc5746SBarry Smith   PetscFunctionBegin;
507827bd09bSSatish Balay   /* to make life easier */
508827bd09bSSatish Balay   nel   = gs->nel;
509827bd09bSSatish Balay   elms  = gs->elms;
510827bd09bSSatish Balay   level = gs->level;
511827bd09bSSatish Balay 
512b1c944f5SJed Brown   /* det #bytes needed for processor bit masks and init w/mask cor. to PCTFS_my_id */
513ca8e9878SJed Brown   p_mask = (PetscInt*) malloc(p_mask_size=PCTFS_len_bit_mask(PCTFS_num_nodes));
514ca8e9878SJed Brown   ierr   = PCTFS_set_bit_mask(p_mask,p_mask_size,PCTFS_my_id);CHKERRQ(ierr);
515827bd09bSSatish Balay 
516827bd09bSSatish Balay   /* allocate space for masks and info bufs */
51752f87cdaSBarry Smith   gs->nghs       = sh_proc_mask = (PetscInt*) malloc(p_mask_size);
51852f87cdaSBarry Smith   gs->pw_nghs    = pw_sh_proc_mask = (PetscInt*) malloc(p_mask_size);
519827bd09bSSatish Balay   gs->ngh_buf_sz = ngh_buf_size = p_mask_size*nel;
52052f87cdaSBarry Smith   t_mask         = (PetscInt*) malloc(p_mask_size);
52152f87cdaSBarry Smith   gs->ngh_buf    = ngh_buf = (PetscInt*) malloc(ngh_buf_size);
522827bd09bSSatish Balay 
523827bd09bSSatish Balay   /* comm buffer size ... memory usage bounded by ~2*msg_buf */
524827bd09bSSatish Balay   /* had thought I could exploit rendezvous threshold */
525827bd09bSSatish Balay 
526827bd09bSSatish Balay   /* default is one pass */
527827bd09bSSatish Balay   per_load      = negl  = gs->negl;
528827bd09bSSatish Balay   gs->num_loads = num_loads = 1;
529827bd09bSSatish Balay   i             = p_mask_size*negl;
530827bd09bSSatish Balay 
531827bd09bSSatish Balay   /* possible overflow on buffer size */
532827bd09bSSatish Balay   /* overflow hack                    */
5332fa5cd67SKarl Rupp   if (i<0) i=INT_MAX;
534827bd09bSSatish Balay 
53539945688SSatish Balay   buf_size = PetscMin(msg_buf,i);
536827bd09bSSatish Balay 
537827bd09bSSatish Balay   /* can we do it? */
538e32f2f54SBarry Smith   if (p_mask_size>buf_size) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"get_ngh_buf() :: buf<pms :: %d>%d\n",p_mask_size,buf_size);
539827bd09bSSatish Balay 
540b1c944f5SJed Brown   /* get PCTFS_giop buf space ... make *only* one malloc */
54152f87cdaSBarry Smith   buf1 = (PetscInt*) malloc(buf_size<<1);
542827bd09bSSatish Balay 
543827bd09bSSatish Balay   /* more than one gior exchange needed? */
544db4deed7SKarl Rupp   if (buf_size!=i) {
545827bd09bSSatish Balay     per_load      = buf_size/p_mask_size;
546827bd09bSSatish Balay     buf_size      = per_load*p_mask_size;
547827bd09bSSatish Balay     gs->num_loads = num_loads = negl/per_load + (negl%per_load>0);
548827bd09bSSatish Balay   }
549827bd09bSSatish Balay 
550827bd09bSSatish Balay 
551827bd09bSSatish Balay   /* convert buf sizes from #bytes to #ints - 32 bit only! */
552a501084fSBarry Smith   p_mask_size/=sizeof(PetscInt); ngh_buf_size/=sizeof(PetscInt); buf_size/=sizeof(PetscInt);
553827bd09bSSatish Balay 
554b1c944f5SJed Brown   /* find PCTFS_giop work space */
555827bd09bSSatish Balay   buf2 = buf1+buf_size;
556827bd09bSSatish Balay 
557827bd09bSSatish Balay   /* hold #ints needed for processor masks */
558827bd09bSSatish Balay   gs->mask_sz=p_mask_size;
559827bd09bSSatish Balay 
560827bd09bSSatish Balay   /* init buffers */
561ca8e9878SJed Brown   ierr = PCTFS_ivec_zero(sh_proc_mask,p_mask_size);CHKERRQ(ierr);
562ca8e9878SJed Brown   ierr = PCTFS_ivec_zero(pw_sh_proc_mask,p_mask_size);CHKERRQ(ierr);
563ca8e9878SJed Brown   ierr = PCTFS_ivec_zero(ngh_buf,ngh_buf_size);CHKERRQ(ierr);
564827bd09bSSatish Balay 
565827bd09bSSatish Balay   /* HACK reset tree info */
566827bd09bSSatish Balay   tree_buf    = NULL;
567827bd09bSSatish Balay   tree_buf_sz = ntree = 0;
568827bd09bSSatish Balay 
569827bd09bSSatish Balay   /* ok do it */
570db4deed7SKarl Rupp   for (ptr1=ngh_buf,ptr2=elms,end=gs->gl_min,or_ct=i=0; or_ct<num_loads; or_ct++) {
571827bd09bSSatish Balay     /* identity for bitwise or is 000...000 */
572ca8e9878SJed Brown     PCTFS_ivec_zero(buf1,buf_size);
573827bd09bSSatish Balay 
574827bd09bSSatish Balay     /* load msg buffer */
575db4deed7SKarl Rupp     for (start=end,end+=per_load,i_start=i; (offset=*ptr2)<end; i++, ptr2++) {
576827bd09bSSatish Balay       offset = (offset-start)*p_mask_size;
577ca8e9878SJed Brown       PCTFS_ivec_copy(buf1+offset,p_mask,p_mask_size);
578827bd09bSSatish Balay     }
579827bd09bSSatish Balay 
580827bd09bSSatish Balay     /* GLOBAL: pass buffer */
581b1c944f5SJed Brown     ierr = PCTFS_giop(buf1,buf2,buf_size,&oper);CHKERRQ(ierr);
582827bd09bSSatish Balay 
583827bd09bSSatish Balay 
584827bd09bSSatish Balay     /* unload buffer into ngh_buf */
585827bd09bSSatish Balay     ptr2=(elms+i_start);
586db4deed7SKarl Rupp     for (ptr3=buf1,j=start; j<end; ptr3+=p_mask_size,j++) {
587827bd09bSSatish Balay       /* I own it ... may have to pairwise it */
588db4deed7SKarl Rupp       if (j==*ptr2) {
589827bd09bSSatish Balay         /* do i share it w/anyone? */
590ca8e9878SJed Brown         ct1 = PCTFS_ct_bits((char*)ptr3,p_mask_size*sizeof(PetscInt));
591827bd09bSSatish Balay         /* guess not */
592db4deed7SKarl Rupp         if (ct1<2) { ptr2++; ptr1+=p_mask_size; continue; }
593827bd09bSSatish Balay 
594827bd09bSSatish Balay         /* i do ... so keep info and turn off my bit */
595ca8e9878SJed Brown         PCTFS_ivec_copy(ptr1,ptr3,p_mask_size);
596ca8e9878SJed Brown         ierr = PCTFS_ivec_xor(ptr1,p_mask,p_mask_size);CHKERRQ(ierr);
597ca8e9878SJed Brown         ierr = PCTFS_ivec_or(sh_proc_mask,ptr1,p_mask_size);CHKERRQ(ierr);
598827bd09bSSatish Balay 
599827bd09bSSatish Balay         /* is it to be done pairwise? */
600db4deed7SKarl Rupp         if (--ct1<=level) {
601827bd09bSSatish Balay           npw++;
602827bd09bSSatish Balay 
603827bd09bSSatish Balay           /* turn on high bit to indicate pw need to process */
604827bd09bSSatish Balay           *ptr2++ |= TOP_BIT;
605ca8e9878SJed Brown           ierr     = PCTFS_ivec_or(pw_sh_proc_mask,ptr1,p_mask_size);CHKERRQ(ierr);
606827bd09bSSatish Balay           ptr1    += p_mask_size;
607827bd09bSSatish Balay           continue;
608827bd09bSSatish Balay         }
609827bd09bSSatish Balay 
610827bd09bSSatish Balay         /* get set for next and note that I have a tree contribution */
611827bd09bSSatish Balay         /* could save exact elm index for tree here -> save a search */
612827bd09bSSatish Balay         ptr2++; ptr1+=p_mask_size; ntree_map++;
613db4deed7SKarl Rupp       } else { /* i don't but still might be involved in tree */
614827bd09bSSatish Balay 
615827bd09bSSatish Balay         /* shared by how many? */
616ca8e9878SJed Brown         ct1 = PCTFS_ct_bits((char*)ptr3,p_mask_size*sizeof(PetscInt));
617827bd09bSSatish Balay 
618827bd09bSSatish Balay         /* none! */
619f1ed62a8SBarry Smith         if (ct1<2) continue;
620827bd09bSSatish Balay 
621827bd09bSSatish Balay         /* is it going to be done pairwise? but not by me of course!*/
622f1ed62a8SBarry Smith         if (--ct1<=level) continue;
623827bd09bSSatish Balay       }
624827bd09bSSatish Balay       /* LATER we're going to have to process it NOW */
625827bd09bSSatish Balay       /* nope ... tree it */
626f1ed62a8SBarry Smith       ierr = place_in_tree(j);CHKERRQ(ierr);
627827bd09bSSatish Balay     }
628827bd09bSSatish Balay   }
629827bd09bSSatish Balay 
630a501084fSBarry Smith   free((void*)t_mask);
631a501084fSBarry Smith   free((void*)buf1);
632827bd09bSSatish Balay 
633827bd09bSSatish Balay   gs->len_pw_list = npw;
634ca8e9878SJed Brown   gs->num_nghs    = PCTFS_ct_bits((char*)sh_proc_mask,p_mask_size*sizeof(PetscInt));
635827bd09bSSatish Balay 
636827bd09bSSatish Balay   /* expand from bit mask list to int list and save ngh list */
63752f87cdaSBarry Smith   gs->nghs = (PetscInt*) malloc(gs->num_nghs * sizeof(PetscInt));
638ca8e9878SJed Brown   PCTFS_bm_to_proc((char*)sh_proc_mask,p_mask_size*sizeof(PetscInt),gs->nghs);
639827bd09bSSatish Balay 
640ca8e9878SJed Brown   gs->num_pw_nghs = PCTFS_ct_bits((char*)pw_sh_proc_mask,p_mask_size*sizeof(PetscInt));
641827bd09bSSatish Balay 
642827bd09bSSatish Balay   oper         = GL_MAX;
643827bd09bSSatish Balay   ct1          = gs->num_nghs;
644b1c944f5SJed Brown   ierr         = PCTFS_giop(&ct1,&ct2,1,&oper);CHKERRQ(ierr);
645827bd09bSSatish Balay   gs->max_nghs = ct1;
646827bd09bSSatish Balay 
647827bd09bSSatish Balay   gs->tree_map_sz  = ntree_map;
648827bd09bSSatish Balay   gs->max_left_over=ntree;
649827bd09bSSatish Balay 
650a501084fSBarry Smith   free((void*)p_mask);
651a501084fSBarry Smith   free((void*)sh_proc_mask);
6523fdc5746SBarry Smith   PetscFunctionReturn(0);
653827bd09bSSatish Balay }
654827bd09bSSatish Balay 
655f1ed62a8SBarry Smith /******************************************************************************/
656ca8e9878SJed Brown static PetscErrorCode set_pairwise(PCTFS_gs_id *gs)
657827bd09bSSatish Balay {
65852f87cdaSBarry Smith   PetscInt       i, j;
65952f87cdaSBarry Smith   PetscInt       p_mask_size;
66052f87cdaSBarry Smith   PetscInt       *p_mask, *sh_proc_mask, *tmp_proc_mask;
66152f87cdaSBarry Smith   PetscInt       *ngh_buf, *buf2;
66252f87cdaSBarry Smith   PetscInt       offset;
66352f87cdaSBarry Smith   PetscInt       *msg_list, *msg_size, **msg_nodes, nprs;
66452f87cdaSBarry Smith   PetscInt       *pairwise_elm_list, len_pair_list=0;
66552f87cdaSBarry Smith   PetscInt       *iptr, t1, i_start, nel, *elms;
66652f87cdaSBarry Smith   PetscInt       ct;
667f1ed62a8SBarry Smith   PetscErrorCode ierr;
668827bd09bSSatish Balay 
6693fdc5746SBarry Smith   PetscFunctionBegin;
670827bd09bSSatish Balay   /* to make life easier */
671827bd09bSSatish Balay   nel          = gs->nel;
672827bd09bSSatish Balay   elms         = gs->elms;
673827bd09bSSatish Balay   ngh_buf      = gs->ngh_buf;
674827bd09bSSatish Balay   sh_proc_mask = gs->pw_nghs;
675827bd09bSSatish Balay 
676827bd09bSSatish Balay   /* need a few temp masks */
677ca8e9878SJed Brown   p_mask_size   = PCTFS_len_bit_mask(PCTFS_num_nodes);
67852f87cdaSBarry Smith   p_mask        = (PetscInt*) malloc(p_mask_size);
67952f87cdaSBarry Smith   tmp_proc_mask = (PetscInt*) malloc(p_mask_size);
680827bd09bSSatish Balay 
681b1c944f5SJed Brown   /* set mask to my PCTFS_my_id's bit mask */
682ca8e9878SJed Brown   ierr = PCTFS_set_bit_mask(p_mask,p_mask_size,PCTFS_my_id);CHKERRQ(ierr);
683827bd09bSSatish Balay 
684a501084fSBarry Smith   p_mask_size /= sizeof(PetscInt);
685827bd09bSSatish Balay 
686827bd09bSSatish Balay   len_pair_list   = gs->len_pw_list;
68752f87cdaSBarry Smith   gs->pw_elm_list = pairwise_elm_list=(PetscInt*)malloc((len_pair_list+1)*sizeof(PetscInt));
688827bd09bSSatish Balay 
689827bd09bSSatish Balay   /* how many processors (nghs) do we have to exchange with? */
690ca8e9878SJed Brown   nprs = gs->num_pairs = PCTFS_ct_bits((char*)sh_proc_mask,p_mask_size*sizeof(PetscInt));
691827bd09bSSatish Balay 
692827bd09bSSatish Balay 
693ca8e9878SJed Brown   /* allocate space for PCTFS_gs_gop() info */
69452f87cdaSBarry Smith   gs->pair_list = msg_list  = (PetscInt*)  malloc(sizeof(PetscInt)*nprs);
69552f87cdaSBarry Smith   gs->msg_sizes = msg_size  = (PetscInt*)  malloc(sizeof(PetscInt)*nprs);
69652f87cdaSBarry Smith   gs->node_list = msg_nodes = (PetscInt**) malloc(sizeof(PetscInt*)*(nprs+1));
697827bd09bSSatish Balay 
698827bd09bSSatish Balay   /* init msg_size list */
699ca8e9878SJed Brown   ierr = PCTFS_ivec_zero(msg_size,nprs);CHKERRQ(ierr);
700827bd09bSSatish Balay 
701827bd09bSSatish Balay   /* expand from bit mask list to int list */
702ca8e9878SJed Brown   ierr = PCTFS_bm_to_proc((char*)sh_proc_mask,p_mask_size*sizeof(PetscInt),msg_list);CHKERRQ(ierr);
703827bd09bSSatish Balay 
704827bd09bSSatish Balay   /* keep list of elements being handled pairwise */
705db4deed7SKarl Rupp   for (i=j=0; i<nel; i++) {
706db4deed7SKarl Rupp     if (elms[i] & TOP_BIT) { elms[i] ^= TOP_BIT; pairwise_elm_list[j++] = i; }
707827bd09bSSatish Balay   }
708827bd09bSSatish Balay   pairwise_elm_list[j] = -1;
709827bd09bSSatish Balay 
710a501084fSBarry Smith   gs->msg_ids_out       = (MPI_Request*)  malloc(sizeof(MPI_Request)*(nprs+1));
711827bd09bSSatish Balay   gs->msg_ids_out[nprs] = MPI_REQUEST_NULL;
712a501084fSBarry Smith   gs->msg_ids_in        = (MPI_Request*)  malloc(sizeof(MPI_Request)*(nprs+1));
713827bd09bSSatish Balay   gs->msg_ids_in[nprs]  = MPI_REQUEST_NULL;
714a501084fSBarry Smith   gs->pw_vals           = (PetscScalar*) malloc(sizeof(PetscScalar)*len_pair_list*vec_sz);
715827bd09bSSatish Balay 
716827bd09bSSatish Balay   /* find who goes to each processor */
717db4deed7SKarl Rupp   for (i_start=i=0; i<nprs; i++) {
718827bd09bSSatish Balay     /* processor i's mask */
719ca8e9878SJed Brown     ierr = PCTFS_set_bit_mask(p_mask,p_mask_size*sizeof(PetscInt),msg_list[i]);CHKERRQ(ierr);
720827bd09bSSatish Balay 
721827bd09bSSatish Balay     /* det # going to processor i */
722db4deed7SKarl Rupp     for (ct=j=0; j<len_pair_list; j++) {
723827bd09bSSatish Balay       buf2 = ngh_buf+(pairwise_elm_list[j]*p_mask_size);
724ca8e9878SJed Brown       ierr = PCTFS_ivec_and3(tmp_proc_mask,p_mask,buf2,p_mask_size);CHKERRQ(ierr);
7252fa5cd67SKarl Rupp       if (PCTFS_ct_bits((char*)tmp_proc_mask,p_mask_size*sizeof(PetscInt))) ct++;
726827bd09bSSatish Balay     }
727827bd09bSSatish Balay     msg_size[i] = ct;
72839945688SSatish Balay     i_start     = PetscMax(i_start,ct);
729827bd09bSSatish Balay 
730827bd09bSSatish Balay     /*space to hold nodes in message to first neighbor */
73152f87cdaSBarry Smith     msg_nodes[i] = iptr = (PetscInt*) malloc(sizeof(PetscInt)*(ct+1));
732827bd09bSSatish Balay 
733db4deed7SKarl Rupp     for (j=0;j<len_pair_list;j++) {
734827bd09bSSatish Balay       buf2 = ngh_buf+(pairwise_elm_list[j]*p_mask_size);
735ca8e9878SJed Brown       ierr = PCTFS_ivec_and3(tmp_proc_mask,p_mask,buf2,p_mask_size);CHKERRQ(ierr);
7362fa5cd67SKarl Rupp       if (PCTFS_ct_bits((char*)tmp_proc_mask,p_mask_size*sizeof(PetscInt))) *iptr++ = j;
737827bd09bSSatish Balay     }
738827bd09bSSatish Balay     *iptr = -1;
739827bd09bSSatish Balay   }
740827bd09bSSatish Balay   msg_nodes[nprs] = NULL;
741827bd09bSSatish Balay 
742827bd09bSSatish Balay   j                  = gs->loc_node_pairs=i_start;
743827bd09bSSatish Balay   t1                 = GL_MAX;
744b1c944f5SJed Brown   ierr               = PCTFS_giop(&i_start,&offset,1,&t1);CHKERRQ(ierr);
745827bd09bSSatish Balay   gs->max_node_pairs = i_start;
746827bd09bSSatish Balay 
747827bd09bSSatish Balay   i_start            = j;
748827bd09bSSatish Balay   t1                 = GL_MIN;
749b1c944f5SJed Brown   ierr               = PCTFS_giop(&i_start,&offset,1,&t1);CHKERRQ(ierr);
750827bd09bSSatish Balay   gs->min_node_pairs = i_start;
751827bd09bSSatish Balay 
752827bd09bSSatish Balay   i_start            = j;
753827bd09bSSatish Balay   t1                 = GL_ADD;
754b1c944f5SJed Brown   ierr               = PCTFS_giop(&i_start,&offset,1,&t1);CHKERRQ(ierr);
755b1c944f5SJed Brown   gs->avg_node_pairs = i_start/PCTFS_num_nodes + 1;
756827bd09bSSatish Balay 
757827bd09bSSatish Balay   i_start = nprs;
758827bd09bSSatish Balay   t1      = GL_MAX;
759b1c944f5SJed Brown   PCTFS_giop(&i_start,&offset,1,&t1);
760827bd09bSSatish Balay   gs->max_pairs = i_start;
761827bd09bSSatish Balay 
762827bd09bSSatish Balay 
763827bd09bSSatish Balay   /* remap pairwise in tail of gsi_via_bit_mask() */
764ca8e9878SJed Brown   gs->msg_total = PCTFS_ivec_sum(gs->msg_sizes,nprs);
765a501084fSBarry Smith   gs->out       = (PetscScalar*) malloc(sizeof(PetscScalar)*gs->msg_total*vec_sz);
766a501084fSBarry Smith   gs->in        = (PetscScalar*) malloc(sizeof(PetscScalar)*gs->msg_total*vec_sz);
767827bd09bSSatish Balay 
768827bd09bSSatish Balay   /* reset malloc pool */
769a501084fSBarry Smith   free((void*)p_mask);
770a501084fSBarry Smith   free((void*)tmp_proc_mask);
7713fdc5746SBarry Smith   PetscFunctionReturn(0);
772827bd09bSSatish Balay }
773827bd09bSSatish Balay 
774f1ed62a8SBarry Smith /* to do pruned tree just save ngh buf copy for each one and decode here!
775827bd09bSSatish Balay ******************************************************************************/
776ca8e9878SJed Brown static PetscErrorCode set_tree(PCTFS_gs_id *gs)
777827bd09bSSatish Balay {
77852f87cdaSBarry Smith   PetscInt i, j, n, nel;
77952f87cdaSBarry Smith   PetscInt *iptr_in, *iptr_out, *tree_elms, *elms;
780827bd09bSSatish Balay 
7813fdc5746SBarry Smith   PetscFunctionBegin;
782827bd09bSSatish Balay   /* local work ptrs */
783827bd09bSSatish Balay   elms = gs->elms;
784827bd09bSSatish Balay   nel  = gs->nel;
785827bd09bSSatish Balay 
786827bd09bSSatish Balay   /* how many via tree */
787827bd09bSSatish Balay   gs->tree_nel     = n = ntree;
788827bd09bSSatish Balay   gs->tree_elms    = tree_elms = iptr_in = tree_buf;
789a501084fSBarry Smith   gs->tree_buf     = (PetscScalar*) malloc(sizeof(PetscScalar)*n*vec_sz);
790a501084fSBarry Smith   gs->tree_work    = (PetscScalar*) malloc(sizeof(PetscScalar)*n*vec_sz);
791827bd09bSSatish Balay   j                = gs->tree_map_sz;
79252f87cdaSBarry Smith   gs->tree_map_in  = iptr_in  = (PetscInt*) malloc(sizeof(PetscInt)*(j+1));
79352f87cdaSBarry Smith   gs->tree_map_out = iptr_out = (PetscInt*) malloc(sizeof(PetscInt)*(j+1));
794827bd09bSSatish Balay 
795827bd09bSSatish Balay   /* search the longer of the two lists */
796827bd09bSSatish Balay   /* note ... could save this info in get_ngh_buf and save searches */
797db4deed7SKarl Rupp   if (n<=nel) {
798827bd09bSSatish Balay     /* bijective fct w/remap - search elm list */
799db4deed7SKarl Rupp     for (i=0; i<n; i++) {
800db4deed7SKarl Rupp       if ((j=PCTFS_ivec_binary_search(*tree_elms++,elms,nel))>=0) {*iptr_in++ = j; *iptr_out++ = i;}
801827bd09bSSatish Balay     }
802db4deed7SKarl Rupp   } else {
803db4deed7SKarl Rupp     for (i=0; i<nel; i++) {
804db4deed7SKarl Rupp       if ((j=PCTFS_ivec_binary_search(*elms++,tree_elms,n))>=0) {*iptr_in++ = i; *iptr_out++ = j;}
805827bd09bSSatish Balay     }
806827bd09bSSatish Balay   }
807827bd09bSSatish Balay 
808827bd09bSSatish Balay   /* sentinel */
809827bd09bSSatish Balay   *iptr_in = *iptr_out = -1;
8103fdc5746SBarry Smith   PetscFunctionReturn(0);
811827bd09bSSatish Balay }
812827bd09bSSatish Balay 
813f1ed62a8SBarry Smith /******************************************************************************/
814ca8e9878SJed Brown static PetscErrorCode PCTFS_gs_gop_local_out(PCTFS_gs_id *gs,  PetscScalar *vals)
815827bd09bSSatish Balay {
81652f87cdaSBarry Smith   PetscInt    *num, *map, **reduce;
817a501084fSBarry Smith   PetscScalar tmp;
818827bd09bSSatish Balay 
8193fdc5746SBarry Smith   PetscFunctionBegin;
820827bd09bSSatish Balay   num    = gs->num_gop_local_reduce;
821827bd09bSSatish Balay   reduce = gs->gop_local_reduce;
822db4deed7SKarl Rupp   while ((map = *reduce++)) {
823827bd09bSSatish Balay     /* wall */
824db4deed7SKarl Rupp     if (*num == 2) {
825827bd09bSSatish Balay       num++;
826827bd09bSSatish Balay       vals[map[1]] = vals[map[0]];
827db4deed7SKarl Rupp     } else if (*num == 3) { /* corner shared by three elements */
828827bd09bSSatish Balay       num++;
829827bd09bSSatish Balay       vals[map[2]] = vals[map[1]] = vals[map[0]];
830db4deed7SKarl Rupp     } else if (*num == 4) { /* corner shared by four elements */
831827bd09bSSatish Balay       num++;
832827bd09bSSatish Balay       vals[map[3]] = vals[map[2]] = vals[map[1]] = vals[map[0]];
833db4deed7SKarl Rupp     } else { /* general case ... odd geoms ... 3D*/
834827bd09bSSatish Balay       num++;
835827bd09bSSatish Balay       tmp = *(vals + *map++);
8362fa5cd67SKarl Rupp       while (*map >= 0) *(vals + *map++) = tmp;
837827bd09bSSatish Balay     }
838827bd09bSSatish Balay   }
8393fdc5746SBarry Smith   PetscFunctionReturn(0);
840827bd09bSSatish Balay }
841827bd09bSSatish Balay 
8427b1ae94cSBarry Smith /******************************************************************************/
843ca8e9878SJed Brown static PetscErrorCode PCTFS_gs_gop_local_plus(PCTFS_gs_id *gs,  PetscScalar *vals)
844827bd09bSSatish Balay {
84552f87cdaSBarry Smith   PetscInt    *num, *map, **reduce;
846a501084fSBarry Smith   PetscScalar tmp;
847827bd09bSSatish Balay 
8483fdc5746SBarry Smith   PetscFunctionBegin;
849827bd09bSSatish Balay   num    = gs->num_local_reduce;
850827bd09bSSatish Balay   reduce = gs->local_reduce;
851db4deed7SKarl Rupp   while ((map = *reduce)) {
852827bd09bSSatish Balay     /* wall */
853db4deed7SKarl Rupp     if (*num == 2) {
854827bd09bSSatish Balay       num++; reduce++;
855827bd09bSSatish Balay       vals[map[1]] = vals[map[0]] += vals[map[1]];
856db4deed7SKarl Rupp     } else if (*num == 3) { /* corner shared by three elements */
857827bd09bSSatish Balay       num++; reduce++;
858827bd09bSSatish Balay       vals[map[2]]=vals[map[1]]=vals[map[0]]+=(vals[map[1]]+vals[map[2]]);
859db4deed7SKarl Rupp     } else if (*num == 4) { /* corner shared by four elements */
860827bd09bSSatish Balay       num++; reduce++;
8612fa5cd67SKarl Rupp       vals[map[1]]=vals[map[2]]=vals[map[3]]=vals[map[0]] += (vals[map[1]] + vals[map[2]] + vals[map[3]]);
862db4deed7SKarl Rupp     } else { /* general case ... odd geoms ... 3D*/
863827bd09bSSatish Balay       num++;
864827bd09bSSatish Balay       tmp = 0.0;
8652fa5cd67SKarl Rupp       while (*map >= 0) tmp += *(vals + *map++);
866827bd09bSSatish Balay 
867827bd09bSSatish Balay       map = *reduce++;
8682fa5cd67SKarl Rupp       while (*map >= 0) *(vals + *map++) = tmp;
869827bd09bSSatish Balay     }
870827bd09bSSatish Balay   }
8713fdc5746SBarry Smith   PetscFunctionReturn(0);
872827bd09bSSatish Balay }
873827bd09bSSatish Balay 
8747b1ae94cSBarry Smith /******************************************************************************/
875ca8e9878SJed Brown static PetscErrorCode PCTFS_gs_gop_local_in_plus(PCTFS_gs_id *gs,  PetscScalar *vals)
876827bd09bSSatish Balay {
87752f87cdaSBarry Smith   PetscInt    *num, *map, **reduce;
878a501084fSBarry Smith   PetscScalar *base;
879827bd09bSSatish Balay 
8803fdc5746SBarry Smith   PetscFunctionBegin;
881827bd09bSSatish Balay   num    = gs->num_gop_local_reduce;
882827bd09bSSatish Balay   reduce = gs->gop_local_reduce;
883db4deed7SKarl Rupp   while ((map = *reduce++)) {
884827bd09bSSatish Balay     /* wall */
885db4deed7SKarl Rupp     if (*num == 2) {
886827bd09bSSatish Balay       num++;
887827bd09bSSatish Balay       vals[map[0]] += vals[map[1]];
888db4deed7SKarl Rupp     } else if (*num == 3) { /* corner shared by three elements */
889827bd09bSSatish Balay       num++;
890827bd09bSSatish Balay       vals[map[0]] += (vals[map[1]] + vals[map[2]]);
891db4deed7SKarl Rupp     } else if (*num == 4) { /* corner shared by four elements */
892827bd09bSSatish Balay       num++;
893827bd09bSSatish Balay       vals[map[0]] += (vals[map[1]] + vals[map[2]] + vals[map[3]]);
894db4deed7SKarl Rupp     } else { /* general case ... odd geoms ... 3D*/
895827bd09bSSatish Balay       num++;
896827bd09bSSatish Balay       base = vals + *map++;
8972fa5cd67SKarl Rupp       while (*map >= 0) *base += *(vals + *map++);
898827bd09bSSatish Balay     }
899827bd09bSSatish Balay   }
9003fdc5746SBarry Smith   PetscFunctionReturn(0);
901827bd09bSSatish Balay }
902827bd09bSSatish Balay 
9037b1ae94cSBarry Smith /******************************************************************************/
904ca8e9878SJed Brown PetscErrorCode PCTFS_gs_free(PCTFS_gs_id *gs)
905827bd09bSSatish Balay {
90652f87cdaSBarry Smith   PetscInt       i;
907*6b967228SBarry Smith   PetscErrorCode ierr;
908827bd09bSSatish Balay 
9093fdc5746SBarry Smith   PetscFunctionBegin;
910*6b967228SBarry Smith   ierr = MPI_Comm_free(&gs->PCTFS_gs_comm);CHKERRABORT(PETSC_COMM_WORLD,ierr);
9112fa5cd67SKarl Rupp   if (gs->nghs) free((void*) gs->nghs);
9122fa5cd67SKarl Rupp   if (gs->pw_nghs) free((void*) gs->pw_nghs);
913827bd09bSSatish Balay 
914827bd09bSSatish Balay   /* tree */
9152fa5cd67SKarl Rupp   if (gs->max_left_over) {
9162fa5cd67SKarl Rupp     if (gs->tree_elms) free((void*) gs->tree_elms);
9172fa5cd67SKarl Rupp     if (gs->tree_buf) free((void*) gs->tree_buf);
9182fa5cd67SKarl Rupp     if (gs->tree_work) free((void*) gs->tree_work);
9192fa5cd67SKarl Rupp     if (gs->tree_map_in) free((void*) gs->tree_map_in);
9202fa5cd67SKarl Rupp     if (gs->tree_map_out) free((void*) gs->tree_map_out);
921827bd09bSSatish Balay   }
922827bd09bSSatish Balay 
923827bd09bSSatish Balay   /* pairwise info */
9242fa5cd67SKarl Rupp   if (gs->num_pairs) {
925827bd09bSSatish Balay     /* should be NULL already */
9262fa5cd67SKarl Rupp     if (gs->ngh_buf) free((void*) gs->ngh_buf);
9272fa5cd67SKarl Rupp     if (gs->elms) free((void*) gs->elms);
9282fa5cd67SKarl Rupp     if (gs->local_elms) free((void*) gs->local_elms);
9292fa5cd67SKarl Rupp     if (gs->companion) free((void*) gs->companion);
930827bd09bSSatish Balay 
931827bd09bSSatish Balay     /* only set if pairwise */
9322fa5cd67SKarl Rupp     if (gs->vals) free((void*) gs->vals);
9332fa5cd67SKarl Rupp     if (gs->in) free((void*) gs->in);
9342fa5cd67SKarl Rupp     if (gs->out) free((void*) gs->out);
9352fa5cd67SKarl Rupp     if (gs->msg_ids_in) free((void*) gs->msg_ids_in);
9362fa5cd67SKarl Rupp     if (gs->msg_ids_out) free((void*) gs->msg_ids_out);
9372fa5cd67SKarl Rupp     if (gs->pw_vals) free((void*) gs->pw_vals);
9382fa5cd67SKarl Rupp     if (gs->pw_elm_list) free((void*) gs->pw_elm_list);
939db4deed7SKarl Rupp     if (gs->node_list) {
940db4deed7SKarl Rupp       for (i=0;i<gs->num_pairs;i++) {
941db4deed7SKarl Rupp         if (gs->node_list[i])  {
942db4deed7SKarl Rupp           free((void*) gs->node_list[i]);
943db4deed7SKarl Rupp         }
944db4deed7SKarl Rupp       }
945a501084fSBarry Smith       free((void*) gs->node_list);
946827bd09bSSatish Balay     }
9472fa5cd67SKarl Rupp     if (gs->msg_sizes) free((void*) gs->msg_sizes);
9482fa5cd67SKarl Rupp     if (gs->pair_list) free((void*) gs->pair_list);
949827bd09bSSatish Balay   }
950827bd09bSSatish Balay 
951827bd09bSSatish Balay   /* local info */
952db4deed7SKarl Rupp   if (gs->num_local_total>=0) {
953db4deed7SKarl Rupp     for (i=0;i<gs->num_local_total+1;i++) {
9542fa5cd67SKarl Rupp       if (gs->num_gop_local_reduce[i]) free((void*) gs->gop_local_reduce[i]);
955827bd09bSSatish Balay     }
956827bd09bSSatish Balay   }
957827bd09bSSatish Balay 
958827bd09bSSatish Balay   /* if intersection tree/pairwise and local isn't empty */
9592fa5cd67SKarl Rupp   if (gs->gop_local_reduce) free((void*) gs->gop_local_reduce);
9602fa5cd67SKarl Rupp   if (gs->num_gop_local_reduce) free((void*) gs->num_gop_local_reduce);
961827bd09bSSatish Balay 
962a501084fSBarry Smith   free((void*) gs);
9633fdc5746SBarry Smith   PetscFunctionReturn(0);
964827bd09bSSatish Balay }
965827bd09bSSatish Balay 
9667b1ae94cSBarry Smith /******************************************************************************/
967ca8e9878SJed Brown PetscErrorCode PCTFS_gs_gop_vec(PCTFS_gs_id *gs,  PetscScalar *vals,  const char *op,  PetscInt step)
968827bd09bSSatish Balay {
969d1528f56SBarry Smith   PetscErrorCode ierr;
970d1528f56SBarry Smith 
9713fdc5746SBarry Smith   PetscFunctionBegin;
972827bd09bSSatish Balay   switch (*op) {
973827bd09bSSatish Balay   case '+':
974ca8e9878SJed Brown     PCTFS_gs_gop_vec_plus(gs,vals,step);
975827bd09bSSatish Balay     break;
976827bd09bSSatish Balay   default:
977955c1f14SBarry Smith     ierr = PetscInfo1(0,"PCTFS_gs_gop_vec() :: %c is not a valid op\n",op[0]);CHKERRQ(ierr);
978955c1f14SBarry Smith     ierr = PetscInfo(0,"PCTFS_gs_gop_vec() :: default :: plus\n");CHKERRQ(ierr);
979ca8e9878SJed Brown     PCTFS_gs_gop_vec_plus(gs,vals,step);
980827bd09bSSatish Balay     break;
981827bd09bSSatish Balay   }
9823fdc5746SBarry Smith   PetscFunctionReturn(0);
983827bd09bSSatish Balay }
984827bd09bSSatish Balay 
9857b1ae94cSBarry Smith /******************************************************************************/
986ca8e9878SJed Brown static PetscErrorCode PCTFS_gs_gop_vec_plus(PCTFS_gs_id *gs,  PetscScalar *vals,  PetscInt step)
987827bd09bSSatish Balay {
9883fdc5746SBarry Smith   PetscFunctionBegin;
989ca8e9878SJed Brown   if (!gs) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"PCTFS_gs_gop_vec() passed NULL gs handle!!!");
990827bd09bSSatish Balay 
991827bd09bSSatish Balay   /* local only operations!!! */
9922fa5cd67SKarl Rupp   if (gs->num_local) PCTFS_gs_gop_vec_local_plus(gs,vals,step);
993827bd09bSSatish Balay 
994827bd09bSSatish Balay   /* if intersection tree/pairwise and local isn't empty */
9952fa5cd67SKarl Rupp   if (gs->num_local_gop) {
996ca8e9878SJed Brown     PCTFS_gs_gop_vec_local_in_plus(gs,vals,step);
997827bd09bSSatish Balay 
998827bd09bSSatish Balay     /* pairwise */
9992fa5cd67SKarl Rupp     if (gs->num_pairs) PCTFS_gs_gop_vec_pairwise_plus(gs,vals,step);
1000827bd09bSSatish Balay 
1001827bd09bSSatish Balay     /* tree */
10022fa5cd67SKarl Rupp     else if (gs->max_left_over) PCTFS_gs_gop_vec_tree_plus(gs,vals,step);
1003827bd09bSSatish Balay 
1004ca8e9878SJed Brown     PCTFS_gs_gop_vec_local_out(gs,vals,step);
1005db4deed7SKarl Rupp   } else { /* if intersection tree/pairwise and local is empty */
1006827bd09bSSatish Balay     /* pairwise */
10072fa5cd67SKarl Rupp     if (gs->num_pairs) PCTFS_gs_gop_vec_pairwise_plus(gs,vals,step);
1008827bd09bSSatish Balay 
1009827bd09bSSatish Balay     /* tree */
10102fa5cd67SKarl Rupp     else if (gs->max_left_over) PCTFS_gs_gop_vec_tree_plus(gs,vals,step);
1011827bd09bSSatish Balay   }
10123fdc5746SBarry Smith   PetscFunctionReturn(0);
1013827bd09bSSatish Balay }
1014827bd09bSSatish Balay 
10157b1ae94cSBarry Smith /******************************************************************************/
1016ca8e9878SJed Brown static PetscErrorCode PCTFS_gs_gop_vec_local_plus(PCTFS_gs_id *gs,  PetscScalar *vals, PetscInt step)
1017827bd09bSSatish Balay {
101852f87cdaSBarry Smith   PetscInt    *num, *map, **reduce;
1019a501084fSBarry Smith   PetscScalar *base;
1020827bd09bSSatish Balay 
10213fdc5746SBarry Smith   PetscFunctionBegin;
1022827bd09bSSatish Balay   num    = gs->num_local_reduce;
1023827bd09bSSatish Balay   reduce = gs->local_reduce;
1024db4deed7SKarl Rupp   while ((map = *reduce)) {
1025827bd09bSSatish Balay     base = vals + map[0] * step;
1026827bd09bSSatish Balay 
1027827bd09bSSatish Balay     /* wall */
1028db4deed7SKarl Rupp     if (*num == 2) {
1029827bd09bSSatish Balay       num++; reduce++;
1030ca8e9878SJed Brown       PCTFS_rvec_add (base,vals+map[1]*step,step);
1031ca8e9878SJed Brown       PCTFS_rvec_copy(vals+map[1]*step,base,step);
1032db4deed7SKarl Rupp     } else if (*num == 3) { /* corner shared by three elements */
1033827bd09bSSatish Balay       num++; reduce++;
1034ca8e9878SJed Brown       PCTFS_rvec_add (base,vals+map[1]*step,step);
1035ca8e9878SJed Brown       PCTFS_rvec_add (base,vals+map[2]*step,step);
1036ca8e9878SJed Brown       PCTFS_rvec_copy(vals+map[2]*step,base,step);
1037ca8e9878SJed Brown       PCTFS_rvec_copy(vals+map[1]*step,base,step);
1038db4deed7SKarl Rupp     } else if (*num == 4) { /* corner shared by four elements */
1039827bd09bSSatish Balay       num++; reduce++;
1040ca8e9878SJed Brown       PCTFS_rvec_add (base,vals+map[1]*step,step);
1041ca8e9878SJed Brown       PCTFS_rvec_add (base,vals+map[2]*step,step);
1042ca8e9878SJed Brown       PCTFS_rvec_add (base,vals+map[3]*step,step);
1043ca8e9878SJed Brown       PCTFS_rvec_copy(vals+map[3]*step,base,step);
1044ca8e9878SJed Brown       PCTFS_rvec_copy(vals+map[2]*step,base,step);
1045ca8e9878SJed Brown       PCTFS_rvec_copy(vals+map[1]*step,base,step);
1046db4deed7SKarl Rupp     } else { /* general case ... odd geoms ... 3D */
1047827bd09bSSatish Balay       num++;
10482fa5cd67SKarl Rupp       while (*++map >= 0) PCTFS_rvec_add (base,vals+*map*step,step);
1049827bd09bSSatish Balay 
1050827bd09bSSatish Balay       map = *reduce;
10512fa5cd67SKarl Rupp       while (*++map >= 0) PCTFS_rvec_copy(vals+*map*step,base,step);
1052827bd09bSSatish Balay 
1053827bd09bSSatish Balay       reduce++;
1054827bd09bSSatish Balay     }
1055827bd09bSSatish Balay   }
10563fdc5746SBarry Smith   PetscFunctionReturn(0);
1057827bd09bSSatish Balay }
1058827bd09bSSatish Balay 
10597b1ae94cSBarry Smith /******************************************************************************/
1060ca8e9878SJed Brown static PetscErrorCode PCTFS_gs_gop_vec_local_in_plus(PCTFS_gs_id *gs,  PetscScalar *vals, PetscInt step)
1061827bd09bSSatish Balay {
106252f87cdaSBarry Smith   PetscInt    *num, *map, **reduce;
1063a501084fSBarry Smith   PetscScalar *base;
1064db4deed7SKarl Rupp 
10653fdc5746SBarry Smith   PetscFunctionBegin;
1066827bd09bSSatish Balay   num    = gs->num_gop_local_reduce;
1067827bd09bSSatish Balay   reduce = gs->gop_local_reduce;
1068db4deed7SKarl Rupp   while ((map = *reduce++)) {
1069827bd09bSSatish Balay     base = vals + map[0] * step;
1070827bd09bSSatish Balay 
1071827bd09bSSatish Balay     /* wall */
1072db4deed7SKarl Rupp     if (*num == 2) {
1073827bd09bSSatish Balay       num++;
1074ca8e9878SJed Brown       PCTFS_rvec_add(base,vals+map[1]*step,step);
1075db4deed7SKarl Rupp     } else if (*num == 3) { /* corner shared by three elements */
1076827bd09bSSatish Balay       num++;
1077ca8e9878SJed Brown       PCTFS_rvec_add(base,vals+map[1]*step,step);
1078ca8e9878SJed Brown       PCTFS_rvec_add(base,vals+map[2]*step,step);
1079db4deed7SKarl Rupp     } else if (*num == 4) { /* corner shared by four elements */
1080827bd09bSSatish Balay       num++;
1081ca8e9878SJed Brown       PCTFS_rvec_add(base,vals+map[1]*step,step);
1082ca8e9878SJed Brown       PCTFS_rvec_add(base,vals+map[2]*step,step);
1083ca8e9878SJed Brown       PCTFS_rvec_add(base,vals+map[3]*step,step);
1084db4deed7SKarl Rupp     } else { /* general case ... odd geoms ... 3D*/
1085827bd09bSSatish Balay       num++;
10862fa5cd67SKarl Rupp       while (*++map >= 0) PCTFS_rvec_add(base,vals+*map*step,step);
1087827bd09bSSatish Balay     }
1088827bd09bSSatish Balay   }
10893fdc5746SBarry Smith   PetscFunctionReturn(0);
1090827bd09bSSatish Balay }
1091827bd09bSSatish Balay 
10927b1ae94cSBarry Smith /******************************************************************************/
1093ca8e9878SJed Brown static PetscErrorCode PCTFS_gs_gop_vec_local_out(PCTFS_gs_id *gs,  PetscScalar *vals, PetscInt step)
1094827bd09bSSatish Balay {
109552f87cdaSBarry Smith   PetscInt    *num, *map, **reduce;
1096a501084fSBarry Smith   PetscScalar *base;
1097827bd09bSSatish Balay 
10983fdc5746SBarry Smith   PetscFunctionBegin;
1099827bd09bSSatish Balay   num    = gs->num_gop_local_reduce;
1100827bd09bSSatish Balay   reduce = gs->gop_local_reduce;
1101db4deed7SKarl Rupp   while ((map = *reduce++)) {
1102827bd09bSSatish Balay     base = vals + map[0] * step;
1103827bd09bSSatish Balay 
1104827bd09bSSatish Balay     /* wall */
1105db4deed7SKarl Rupp     if (*num == 2) {
1106827bd09bSSatish Balay       num++;
1107ca8e9878SJed Brown       PCTFS_rvec_copy(vals+map[1]*step,base,step);
1108db4deed7SKarl Rupp     } else if (*num == 3) { /* corner shared by three elements */
1109827bd09bSSatish Balay       num++;
1110ca8e9878SJed Brown       PCTFS_rvec_copy(vals+map[1]*step,base,step);
1111ca8e9878SJed Brown       PCTFS_rvec_copy(vals+map[2]*step,base,step);
1112db4deed7SKarl Rupp     } else if (*num == 4) { /* corner shared by four elements */
1113827bd09bSSatish Balay       num++;
1114ca8e9878SJed Brown       PCTFS_rvec_copy(vals+map[1]*step,base,step);
1115ca8e9878SJed Brown       PCTFS_rvec_copy(vals+map[2]*step,base,step);
1116ca8e9878SJed Brown       PCTFS_rvec_copy(vals+map[3]*step,base,step);
1117db4deed7SKarl Rupp     } else { /* general case ... odd geoms ... 3D*/
1118827bd09bSSatish Balay       num++;
11192fa5cd67SKarl Rupp       while (*++map >= 0) PCTFS_rvec_copy(vals+*map*step,base,step);
1120827bd09bSSatish Balay     }
1121827bd09bSSatish Balay   }
11223fdc5746SBarry Smith   PetscFunctionReturn(0);
1123827bd09bSSatish Balay }
1124827bd09bSSatish Balay 
11257b1ae94cSBarry Smith /******************************************************************************/
1126ca8e9878SJed Brown static PetscErrorCode PCTFS_gs_gop_vec_pairwise_plus(PCTFS_gs_id *gs,  PetscScalar *in_vals, PetscInt step)
1127827bd09bSSatish Balay {
1128a501084fSBarry Smith   PetscScalar    *dptr1, *dptr2, *dptr3, *in1, *in2;
112952f87cdaSBarry Smith   PetscInt       *iptr, *msg_list, *msg_size, **msg_nodes;
113052f87cdaSBarry Smith   PetscInt       *pw, *list, *size, **nodes;
1131827bd09bSSatish Balay   MPI_Request    *msg_ids_in, *msg_ids_out, *ids_in, *ids_out;
1132827bd09bSSatish Balay   MPI_Status     status;
11330805154bSBarry Smith   PetscBLASInt   i1 = 1,dstep;
11343fdc5746SBarry Smith   PetscErrorCode ierr;
1135827bd09bSSatish Balay 
11363fdc5746SBarry Smith   PetscFunctionBegin;
1137a501084fSBarry Smith   /* strip and load s */
1138827bd09bSSatish Balay   msg_list    = list     = gs->pair_list;
1139827bd09bSSatish Balay   msg_size    = size     = gs->msg_sizes;
1140827bd09bSSatish Balay   msg_nodes   = nodes    = gs->node_list;
1141827bd09bSSatish Balay   iptr        = pw       = gs->pw_elm_list;
1142827bd09bSSatish Balay   dptr1       = dptr3    = gs->pw_vals;
1143827bd09bSSatish Balay   msg_ids_in  = ids_in   = gs->msg_ids_in;
1144827bd09bSSatish Balay   msg_ids_out = ids_out  = gs->msg_ids_out;
1145827bd09bSSatish Balay   dptr2                  = gs->out;
1146827bd09bSSatish Balay   in1=in2                = gs->in;
1147827bd09bSSatish Balay 
1148827bd09bSSatish Balay   /* post the receives */
1149827bd09bSSatish Balay   /*  msg_nodes=nodes; */
1150db4deed7SKarl Rupp   do {
1151827bd09bSSatish Balay     /* Should MPI_ANY_SOURCE be replaced by *list ? In that case do the
1152827bd09bSSatish Balay         second one *list and do list++ afterwards */
1153ca8e9878SJed Brown     ierr = MPI_Irecv(in1, *size *step, MPIU_SCALAR, MPI_ANY_SOURCE, MSGTAG1 + *list, gs->PCTFS_gs_comm, msg_ids_in);CHKERRQ(ierr);
11549182e22cSBarry Smith     list++;msg_ids_in++;
1155827bd09bSSatish Balay     in1 += *size++ *step;
11562fa5cd67SKarl Rupp   } while (*++msg_nodes);
1157827bd09bSSatish Balay   msg_nodes=nodes;
1158827bd09bSSatish Balay 
1159827bd09bSSatish Balay   /* load gs values into in out gs buffers */
1160db4deed7SKarl Rupp   while (*iptr >= 0) {
1161ca8e9878SJed Brown     PCTFS_rvec_copy(dptr3,in_vals + *iptr*step,step);
1162827bd09bSSatish Balay     dptr3+=step;
1163827bd09bSSatish Balay     iptr++;
1164827bd09bSSatish Balay   }
1165827bd09bSSatish Balay 
1166827bd09bSSatish Balay   /* load out buffers and post the sends */
1167db4deed7SKarl Rupp   while ((iptr = *msg_nodes++)) {
1168827bd09bSSatish Balay     dptr3 = dptr2;
1169db4deed7SKarl Rupp     while (*iptr >= 0) {
1170ca8e9878SJed Brown       PCTFS_rvec_copy(dptr2,dptr1 + *iptr*step,step);
1171827bd09bSSatish Balay       dptr2+=step;
1172827bd09bSSatish Balay       iptr++;
1173827bd09bSSatish Balay     }
1174ca8e9878SJed Brown     ierr = MPI_Isend(dptr3, *msg_size *step, MPIU_SCALAR, *msg_list, MSGTAG1+PCTFS_my_id, gs->PCTFS_gs_comm, msg_ids_out);CHKERRQ(ierr);
11759182e22cSBarry Smith     msg_size++; msg_list++;msg_ids_out++;
1176827bd09bSSatish Balay   }
1177827bd09bSSatish Balay 
1178827bd09bSSatish Balay   /* tree */
11792fa5cd67SKarl Rupp   if (gs->max_left_over) PCTFS_gs_gop_vec_tree_plus(gs,in_vals,step);
1180827bd09bSSatish Balay 
1181827bd09bSSatish Balay   /* process the received data */
1182827bd09bSSatish Balay   msg_nodes=nodes;
1183a501084fSBarry Smith   while ((iptr = *nodes++)) {
1184a501084fSBarry Smith     PetscScalar d1 = 1.0;
1185db4deed7SKarl Rupp 
1186827bd09bSSatish Balay     /* Should I check the return value of MPI_Wait() or status? */
1187827bd09bSSatish Balay     /* Can this loop be replaced by a call to MPI_Waitall()? */
11889182e22cSBarry Smith     ierr = MPI_Wait(ids_in, &status);CHKERRQ(ierr);
11899182e22cSBarry Smith     ids_in++;
1190a501084fSBarry Smith     while (*iptr >= 0) {
1191c5df96a5SBarry Smith       ierr = PetscBLASIntCast(step,&dstep);CHKERRQ(ierr);
11928b83055fSJed Brown       PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&dstep,&d1,in2,&i1,dptr1 + *iptr*step,&i1));
1193827bd09bSSatish Balay       in2+=step;
1194827bd09bSSatish Balay       iptr++;
1195827bd09bSSatish Balay     }
1196827bd09bSSatish Balay   }
1197827bd09bSSatish Balay 
1198827bd09bSSatish Balay   /* replace vals */
1199db4deed7SKarl Rupp   while (*pw >= 0) {
1200ca8e9878SJed Brown     PCTFS_rvec_copy(in_vals + *pw*step,dptr1,step);
1201827bd09bSSatish Balay     dptr1+=step;
1202827bd09bSSatish Balay     pw++;
1203827bd09bSSatish Balay   }
1204827bd09bSSatish Balay 
1205827bd09bSSatish Balay   /* clear isend message handles */
1206827bd09bSSatish Balay   /* This changed for clarity though it could be the same */
1207db4deed7SKarl Rupp 
1208827bd09bSSatish Balay   /* Should I check the return value of MPI_Wait() or status? */
1209827bd09bSSatish Balay   /* Can this loop be replaced by a call to MPI_Waitall()? */
12102fa5cd67SKarl Rupp   while (*msg_nodes++) {
12112fa5cd67SKarl Rupp     ierr = MPI_Wait(ids_out, &status);CHKERRQ(ierr);
12122fa5cd67SKarl Rupp     ids_out++;
12132fa5cd67SKarl Rupp   }
12143fdc5746SBarry Smith   PetscFunctionReturn(0);
1215827bd09bSSatish Balay }
1216827bd09bSSatish Balay 
12177b1ae94cSBarry Smith /******************************************************************************/
1218ca8e9878SJed Brown static PetscErrorCode PCTFS_gs_gop_vec_tree_plus(PCTFS_gs_id *gs,  PetscScalar *vals,  PetscInt step)
1219827bd09bSSatish Balay {
122052f87cdaSBarry Smith   PetscInt       size, *in, *out;
1221a501084fSBarry Smith   PetscScalar    *buf, *work;
122252f87cdaSBarry Smith   PetscInt       op[] = {GL_ADD,0};
1223a501084fSBarry Smith   PetscBLASInt   i1   = 1;
1224c5df96a5SBarry Smith   PetscErrorCode ierr;
1225c5df96a5SBarry Smith   PetscBLASInt   dstep;
1226827bd09bSSatish Balay 
12273fdc5746SBarry Smith   PetscFunctionBegin;
1228827bd09bSSatish Balay   /* copy over to local variables */
1229827bd09bSSatish Balay   in   = gs->tree_map_in;
1230827bd09bSSatish Balay   out  = gs->tree_map_out;
1231827bd09bSSatish Balay   buf  = gs->tree_buf;
1232827bd09bSSatish Balay   work = gs->tree_work;
1233827bd09bSSatish Balay   size = gs->tree_nel*step;
1234827bd09bSSatish Balay 
1235827bd09bSSatish Balay   /* zero out collection buffer */
1236ca8e9878SJed Brown   PCTFS_rvec_zero(buf,size);
1237827bd09bSSatish Balay 
1238827bd09bSSatish Balay 
1239827bd09bSSatish Balay   /* copy over my contributions */
1240db4deed7SKarl Rupp   while (*in >= 0) {
1241c5df96a5SBarry Smith     ierr = PetscBLASIntCast(step,&dstep);CHKERRQ(ierr);
12428b83055fSJed Brown     PetscStackCallBLAS("BLAScopy",BLAScopy_(&dstep,vals + *in++ * step,&i1,buf + *out++ * step,&i1));
1243827bd09bSSatish Balay   }
1244827bd09bSSatish Balay 
1245827bd09bSSatish Balay   /* perform fan in/out on full buffer */
1246b1c944f5SJed Brown   /* must change PCTFS_grop to handle the blas */
1247b1c944f5SJed Brown   PCTFS_grop(buf,work,size,op);
1248827bd09bSSatish Balay 
1249827bd09bSSatish Balay   /* reset */
1250827bd09bSSatish Balay   in  = gs->tree_map_in;
1251827bd09bSSatish Balay   out = gs->tree_map_out;
1252827bd09bSSatish Balay 
1253827bd09bSSatish Balay   /* get the portion of the results I need */
1254db4deed7SKarl Rupp   while (*in >= 0) {
1255c5df96a5SBarry Smith     ierr = PetscBLASIntCast(step,&dstep);CHKERRQ(ierr);
12568b83055fSJed Brown     PetscStackCallBLAS("BLAScopy",BLAScopy_(&dstep,buf + *out++ * step,&i1,vals + *in++ * step,&i1));
1257827bd09bSSatish Balay   }
12583fdc5746SBarry Smith   PetscFunctionReturn(0);
1259827bd09bSSatish Balay }
1260827bd09bSSatish Balay 
12617b1ae94cSBarry Smith /******************************************************************************/
1262ca8e9878SJed Brown PetscErrorCode PCTFS_gs_gop_hc(PCTFS_gs_id *gs,  PetscScalar *vals,  const char *op,  PetscInt dim)
1263827bd09bSSatish Balay {
1264d1528f56SBarry Smith   PetscErrorCode ierr;
1265d1528f56SBarry Smith 
12663fdc5746SBarry Smith   PetscFunctionBegin;
1267827bd09bSSatish Balay   switch (*op) {
1268827bd09bSSatish Balay   case '+':
1269ca8e9878SJed Brown     PCTFS_gs_gop_plus_hc(gs,vals,dim);
1270827bd09bSSatish Balay     break;
1271827bd09bSSatish Balay   default:
1272955c1f14SBarry Smith     ierr = PetscInfo1(0,"PCTFS_gs_gop_hc() :: %c is not a valid op\n",op[0]);CHKERRQ(ierr);
1273ca8e9878SJed Brown     ierr = PetscInfo(0,"PCTFS_gs_gop_hc() :: default :: plus\n");CHKERRQ(ierr);
1274ca8e9878SJed Brown     PCTFS_gs_gop_plus_hc(gs,vals,dim);
1275827bd09bSSatish Balay     break;
1276827bd09bSSatish Balay   }
12773fdc5746SBarry Smith   PetscFunctionReturn(0);
1278827bd09bSSatish Balay }
1279827bd09bSSatish Balay 
12807b1ae94cSBarry Smith /******************************************************************************/
1281ca8e9878SJed Brown static PetscErrorCode PCTFS_gs_gop_plus_hc(PCTFS_gs_id *gs,  PetscScalar *vals, PetscInt dim)
1282827bd09bSSatish Balay {
12833fdc5746SBarry Smith   PetscFunctionBegin;
1284827bd09bSSatish Balay   /* if there's nothing to do return */
12852fa5cd67SKarl Rupp   if (dim<=0) PetscFunctionReturn(0);
1286827bd09bSSatish Balay 
1287827bd09bSSatish Balay   /* can't do more dimensions then exist */
1288b1c944f5SJed Brown   dim = PetscMin(dim,PCTFS_i_log2_num_nodes);
1289827bd09bSSatish Balay 
1290827bd09bSSatish Balay   /* local only operations!!! */
12912fa5cd67SKarl Rupp   if (gs->num_local) PCTFS_gs_gop_local_plus(gs,vals);
1292827bd09bSSatish Balay 
1293827bd09bSSatish Balay   /* if intersection tree/pairwise and local isn't empty */
1294db4deed7SKarl Rupp   if (gs->num_local_gop) {
1295ca8e9878SJed Brown     PCTFS_gs_gop_local_in_plus(gs,vals);
1296827bd09bSSatish Balay 
1297827bd09bSSatish Balay     /* pairwise will do tree inside ... */
12982fa5cd67SKarl Rupp     if (gs->num_pairs) PCTFS_gs_gop_pairwise_plus_hc(gs,vals,dim); /* tree only */
12992fa5cd67SKarl Rupp     else if (gs->max_left_over) PCTFS_gs_gop_tree_plus_hc(gs,vals,dim);
1300827bd09bSSatish Balay 
1301ca8e9878SJed Brown     PCTFS_gs_gop_local_out(gs,vals);
1302db4deed7SKarl Rupp   } else { /* if intersection tree/pairwise and local is empty */
1303827bd09bSSatish Balay     /* pairwise will do tree inside */
13042fa5cd67SKarl Rupp     if (gs->num_pairs) PCTFS_gs_gop_pairwise_plus_hc(gs,vals,dim); /* tree */
13052fa5cd67SKarl Rupp     else if (gs->max_left_over) PCTFS_gs_gop_tree_plus_hc(gs,vals,dim);
1306827bd09bSSatish Balay   }
13073fdc5746SBarry Smith   PetscFunctionReturn(0);
1308827bd09bSSatish Balay }
1309827bd09bSSatish Balay 
13107b1ae94cSBarry Smith /******************************************************************************/
1311ca8e9878SJed Brown static PetscErrorCode PCTFS_gs_gop_pairwise_plus_hc(PCTFS_gs_id *gs,  PetscScalar *in_vals, PetscInt dim)
1312827bd09bSSatish Balay {
1313a501084fSBarry Smith   PetscScalar    *dptr1, *dptr2, *dptr3, *in1, *in2;
131452f87cdaSBarry Smith   PetscInt       *iptr, *msg_list, *msg_size, **msg_nodes;
131552f87cdaSBarry Smith   PetscInt       *pw, *list, *size, **nodes;
1316827bd09bSSatish Balay   MPI_Request    *msg_ids_in, *msg_ids_out, *ids_in, *ids_out;
1317827bd09bSSatish Balay   MPI_Status     status;
131852f87cdaSBarry Smith   PetscInt       i, mask=1;
13193fdc5746SBarry Smith   PetscErrorCode ierr;
1320827bd09bSSatish Balay 
13213fdc5746SBarry Smith   PetscFunctionBegin;
1322db4deed7SKarl Rupp   for (i=1; i<dim; i++) { mask<<=1; mask++; }
1323827bd09bSSatish Balay 
1324a501084fSBarry Smith   /* strip and load s */
1325827bd09bSSatish Balay   msg_list    = list     = gs->pair_list;
1326827bd09bSSatish Balay   msg_size    = size     = gs->msg_sizes;
1327827bd09bSSatish Balay   msg_nodes   = nodes    = gs->node_list;
1328827bd09bSSatish Balay   iptr        = pw       = gs->pw_elm_list;
1329827bd09bSSatish Balay   dptr1       = dptr3    = gs->pw_vals;
1330827bd09bSSatish Balay   msg_ids_in  = ids_in   = gs->msg_ids_in;
1331827bd09bSSatish Balay   msg_ids_out = ids_out  = gs->msg_ids_out;
1332827bd09bSSatish Balay   dptr2       = gs->out;
1333827bd09bSSatish Balay   in1         = in2      = gs->in;
1334827bd09bSSatish Balay 
1335827bd09bSSatish Balay   /* post the receives */
1336827bd09bSSatish Balay   /*  msg_nodes=nodes; */
1337db4deed7SKarl Rupp   do {
1338827bd09bSSatish Balay     /* Should MPI_ANY_SOURCE be replaced by *list ? In that case do the
1339827bd09bSSatish Balay         second one *list and do list++ afterwards */
1340db4deed7SKarl Rupp     if ((PCTFS_my_id|mask)==(*list|mask)) {
1341ca8e9878SJed Brown       ierr = MPI_Irecv(in1, *size, MPIU_SCALAR, MPI_ANY_SOURCE, MSGTAG1 + *list, gs->PCTFS_gs_comm, msg_ids_in);CHKERRQ(ierr);
13429182e22cSBarry Smith       list++; msg_ids_in++;in1 += *size++;
1343db4deed7SKarl Rupp     } else { list++; size++; }
13442fa5cd67SKarl Rupp   } while (*++msg_nodes);
1345827bd09bSSatish Balay 
1346827bd09bSSatish Balay   /* load gs values into in out gs buffers */
13472fa5cd67SKarl Rupp   while (*iptr >= 0) *dptr3++ = *(in_vals + *iptr++);
1348827bd09bSSatish Balay 
1349827bd09bSSatish Balay   /* load out buffers and post the sends */
1350827bd09bSSatish Balay   msg_nodes=nodes;
1351827bd09bSSatish Balay   list     = msg_list;
1352db4deed7SKarl Rupp   while ((iptr = *msg_nodes++)) {
1353db4deed7SKarl Rupp     if ((PCTFS_my_id|mask)==(*list|mask)) {
1354827bd09bSSatish Balay       dptr3 = dptr2;
13552fa5cd67SKarl Rupp       while (*iptr >= 0) *dptr2++ = *(dptr1 + *iptr++);
1356827bd09bSSatish Balay       /* CHECK PERSISTENT COMMS MODE FOR ALL THIS STUFF */
1357827bd09bSSatish Balay       /* is msg_ids_out++ correct? */
1358ca8e9878SJed Brown       ierr = MPI_Isend(dptr3, *msg_size, MPIU_SCALAR, *list, MSGTAG1+PCTFS_my_id, gs->PCTFS_gs_comm, msg_ids_out);CHKERRQ(ierr);
13599182e22cSBarry Smith       msg_size++;list++;msg_ids_out++;
1360db4deed7SKarl Rupp     } else {list++; msg_size++;}
1361827bd09bSSatish Balay   }
1362827bd09bSSatish Balay 
1363827bd09bSSatish Balay   /* do the tree while we're waiting */
13642fa5cd67SKarl Rupp   if (gs->max_left_over) PCTFS_gs_gop_tree_plus_hc(gs,in_vals,dim);
1365827bd09bSSatish Balay 
1366827bd09bSSatish Balay   /* process the received data */
1367827bd09bSSatish Balay   msg_nodes=nodes;
1368827bd09bSSatish Balay   list     = msg_list;
1369db4deed7SKarl Rupp   while ((iptr = *nodes++)) {
1370db4deed7SKarl Rupp     if ((PCTFS_my_id|mask)==(*list|mask)) {
1371827bd09bSSatish Balay       /* Should I check the return value of MPI_Wait() or status? */
1372827bd09bSSatish Balay       /* Can this loop be replaced by a call to MPI_Waitall()? */
13739182e22cSBarry Smith       ierr = MPI_Wait(ids_in, &status);CHKERRQ(ierr);
13749182e22cSBarry Smith       ids_in++;
13752fa5cd67SKarl Rupp       while (*iptr >= 0) *(dptr1 + *iptr++) += *in2++;
1376827bd09bSSatish Balay     }
1377827bd09bSSatish Balay     list++;
1378827bd09bSSatish Balay   }
1379827bd09bSSatish Balay 
1380827bd09bSSatish Balay   /* replace vals */
13812fa5cd67SKarl Rupp   while (*pw >= 0) *(in_vals + *pw++) = *dptr1++;
1382827bd09bSSatish Balay 
1383827bd09bSSatish Balay   /* clear isend message handles */
1384827bd09bSSatish Balay   /* This changed for clarity though it could be the same */
1385db4deed7SKarl Rupp   while (*msg_nodes++) {
1386db4deed7SKarl Rupp     if ((PCTFS_my_id|mask)==(*msg_list|mask)) {
1387827bd09bSSatish Balay       /* Should I check the return value of MPI_Wait() or status? */
1388827bd09bSSatish Balay       /* Can this loop be replaced by a call to MPI_Waitall()? */
13899182e22cSBarry Smith       ierr = MPI_Wait(ids_out, &status);CHKERRQ(ierr);
13909182e22cSBarry Smith       ids_out++;
1391827bd09bSSatish Balay     }
1392827bd09bSSatish Balay     msg_list++;
1393827bd09bSSatish Balay   }
13943fdc5746SBarry Smith   PetscFunctionReturn(0);
1395827bd09bSSatish Balay }
1396827bd09bSSatish Balay 
13977b1ae94cSBarry Smith /******************************************************************************/
1398ca8e9878SJed Brown static PetscErrorCode PCTFS_gs_gop_tree_plus_hc(PCTFS_gs_id *gs, PetscScalar *vals, PetscInt dim)
1399827bd09bSSatish Balay {
140052f87cdaSBarry Smith   PetscInt    size;
140152f87cdaSBarry Smith   PetscInt    *in, *out;
1402a501084fSBarry Smith   PetscScalar *buf, *work;
140352f87cdaSBarry Smith   PetscInt    op[] = {GL_ADD,0};
1404827bd09bSSatish Balay 
14053fdc5746SBarry Smith   PetscFunctionBegin;
1406827bd09bSSatish Balay   in   = gs->tree_map_in;
1407827bd09bSSatish Balay   out  = gs->tree_map_out;
1408827bd09bSSatish Balay   buf  = gs->tree_buf;
1409827bd09bSSatish Balay   work = gs->tree_work;
1410827bd09bSSatish Balay   size = gs->tree_nel;
1411827bd09bSSatish Balay 
1412ca8e9878SJed Brown   PCTFS_rvec_zero(buf,size);
1413827bd09bSSatish Balay 
14142fa5cd67SKarl Rupp   while (*in >= 0) *(buf + *out++) = *(vals + *in++);
1415827bd09bSSatish Balay 
1416827bd09bSSatish Balay   in  = gs->tree_map_in;
1417827bd09bSSatish Balay   out = gs->tree_map_out;
1418827bd09bSSatish Balay 
1419b1c944f5SJed Brown   PCTFS_grop_hc(buf,work,size,op,dim);
1420827bd09bSSatish Balay 
14212fa5cd67SKarl Rupp   while (*in >= 0) *(vals + *in++) = *(buf + *out++);
14223fdc5746SBarry Smith   PetscFunctionReturn(0);
1423827bd09bSSatish Balay }
1424827bd09bSSatish Balay 
1425827bd09bSSatish Balay 
1426827bd09bSSatish Balay 
1427