Actual source code: sbaijov.c
1: #define PETSCMAT_DLL
3: /*
4: Routines to compute overlapping regions of a parallel MPI matrix.
5: Used for finding submatrices that were shared across processors.
6: */
7: #include src/mat/impls/sbaij/mpi/mpisbaij.h
8: #include petscbt.h
10: static PetscErrorCode MatIncreaseOverlap_MPISBAIJ_Once(Mat,PetscInt,IS*);
11: static PetscErrorCode MatIncreaseOverlap_MPISBAIJ_Local(Mat,PetscInt*,PetscInt,PetscInt*,PetscBT*);
15: PetscErrorCode MatIncreaseOverlap_MPISBAIJ(Mat C,PetscInt is_max,IS is[],PetscInt ov)
16: {
18: PetscInt i,N=C->N, bs=C->bs;
19: IS *is_new;
22: PetscMalloc(is_max*sizeof(IS),&is_new);
23: /* Convert the indices into block format */
24: ISCompressIndicesGeneral(N,bs,is_max,is,is_new);
25: if (ov < 0){ SETERRQ(PETSC_ERR_ARG_OUTOFRANGE,"Negative overlap specified\n");}
26: for (i=0; i<ov; ++i) {
27: MatIncreaseOverlap_MPISBAIJ_Once(C,is_max,is_new);
28: }
29: for (i=0; i<is_max; i++) {ISDestroy(is[i]);}
30: ISExpandIndicesGeneral(N,bs,is_max,is_new,is);
31: for (i=0; i<is_max; i++) {ISDestroy(is_new[i]);}
32: PetscFree(is_new);
33: return(0);
34: }
36: typedef enum {MINE,OTHER} WhoseOwner;
37: /* data1, odata1 and odata2 are packed in the format (for communication):
38: data[0] = is_max, no of is
39: data[1] = size of is[0]
40: ...
41: data[is_max] = size of is[is_max-1]
42: data[is_max + 1] = data(is[0])
43: ...
44: data[is_max+1+sum(size of is[k]), k=0,...,i-1] = data(is[i])
45: ...
46: data2 is packed in the format (for creating output is[]):
47: data[0] = is_max, no of is
48: data[1] = size of is[0]
49: ...
50: data[is_max] = size of is[is_max-1]
51: data[is_max + 1] = data(is[0])
52: ...
53: data[is_max + 1 + Mbs*i) = data(is[i])
54: ...
55: */
58: static PetscErrorCode MatIncreaseOverlap_MPISBAIJ_Once(Mat C,PetscInt is_max,IS is[])
59: {
60: Mat_MPISBAIJ *c = (Mat_MPISBAIJ*)C->data;
62: PetscMPIInt size,rank,tag1,tag2,*len_s,nrqr,nrqs,*id_r1,*len_r1,flag,len;
63: PetscInt idx,*idx_i,isz,col,*n,*data1,**data1_start,*data2,*data2_i,*data,*data_i,
64: Mbs,i,j,k,*odata1,*odata2,
65: proc_id,**odata2_ptr,*ctable=0,*btable,len_max,len_est;
66: PetscInt proc_end=0,*iwork,len_unused,nodata2;
67: PetscInt ois_max; /* max no of is[] in each of processor */
68: char *t_p;
69: MPI_Comm comm;
70: MPI_Request *s_waits1,*s_waits2,r_req;
71: MPI_Status *s_status,r_status;
72: PetscBT *table; /* mark indices of this processor's is[] */
73: PetscBT table_i;
74: PetscBT otable; /* mark indices of other processors' is[] */
75: PetscInt bs=C->bs,Bn = c->B->n,Bnbs = Bn/bs,*Bowners;
76: IS garray_local,garray_gl;
79: comm = C->comm;
80: size = c->size;
81: rank = c->rank;
82: Mbs = c->Mbs;
84: PetscObjectGetNewTag((PetscObject)C,&tag1);
85: PetscObjectGetNewTag((PetscObject)C,&tag2);
87: /* create tables used in
88: step 1: table[i] - mark c->garray of proc [i]
89: step 3: table[i] - mark indices of is[i] when whose=MINE
90: table[0] - mark incideces of is[] when whose=OTHER */
91: len = PetscMax(is_max, size);
92: len_max = len*sizeof(PetscBT) + (Mbs/PETSC_BITS_PER_BYTE+1)*len*sizeof(char) + 1;
93: PetscMalloc(len_max,&table);
94: t_p = (char *)(table + len);
95: for (i=0; i<len; i++) {
96: table[i] = t_p + (Mbs/PETSC_BITS_PER_BYTE+1)*i;
97: }
99: MPI_Allreduce(&is_max,&ois_max,1,MPIU_INT,MPI_MAX,comm);
100:
101: /* 1. Send this processor's is[] to other processors */
102: /*---------------------------------------------------*/
103: /* allocate spaces */
104: PetscMalloc(is_max*sizeof(PetscInt),&n);
105: len = 0;
106: for (i=0; i<is_max; i++) {
107: ISGetLocalSize(is[i],&n[i]);
108: len += n[i];
109: }
110: if (!len) {
111: is_max = 0;
112: } else {
113: len += 1 + is_max; /* max length of data1 for one processor */
114: }
116:
117: PetscMalloc((size*len+1)*sizeof(PetscInt),&data1);
118: PetscMalloc(size*sizeof(PetscInt*),&data1_start);
119: for (i=0; i<size; i++) data1_start[i] = data1 + i*len;
121: PetscMalloc((size*4+1)*sizeof(PetscInt),&len_s);
122: btable = (PetscInt*)(len_s + size);
123: iwork = btable + size;
124: Bowners = iwork + size;
126: /* gather c->garray from all processors */
127: ISCreateGeneral(comm,Bnbs,c->garray,&garray_local);
128: ISAllGather(garray_local, &garray_gl);
129: ISDestroy(garray_local);
130: MPI_Allgather(&Bnbs,1,MPIU_INT,Bowners+1,1,MPIU_INT,comm);
131: Bowners[0] = 0;
132: for (i=0; i<size; i++) Bowners[i+1] += Bowners[i];
133:
134: if (is_max){
135: /* hash table ctable which maps c->row to proc_id) */
136: PetscMalloc(Mbs*sizeof(PetscInt),&ctable);
137: for (proc_id=0,j=0; proc_id<size; proc_id++) {
138: for (; j<c->rowners[proc_id+1]; j++) {
139: ctable[j] = proc_id;
140: }
141: }
143: /* hash tables marking c->garray */
144: ISGetIndices(garray_gl,&idx_i);
145: for (i=0; i<size; i++){
146: table_i = table[i];
147: PetscBTMemzero(Mbs,table_i);
148: for (j = Bowners[i]; j<Bowners[i+1]; j++){ /* go through B cols of proc[i]*/
149: PetscBTSet(table_i,idx_i[j]);
150: }
151: }
152: ISRestoreIndices(garray_gl,&idx_i);
153: } /* if (is_max) */
154: ISDestroy(garray_gl);
156: /* evaluate communication - mesg to who, length, and buffer space */
157: for (i=0; i<size; i++) len_s[i] = 0;
158:
159: /* header of data1 */
160: for (proc_id=0; proc_id<size; proc_id++){
161: iwork[proc_id] = 0;
162: *data1_start[proc_id] = is_max;
163: data1_start[proc_id]++;
164: for (j=0; j<is_max; j++) {
165: if (proc_id == rank){
166: *data1_start[proc_id] = n[j];
167: } else {
168: *data1_start[proc_id] = 0;
169: }
170: data1_start[proc_id]++;
171: }
172: }
173:
174: for (i=0; i<is_max; i++) {
175: ISGetIndices(is[i],&idx_i);
176: for (j=0; j<n[i]; j++){
177: idx = idx_i[j];
178: *data1_start[rank] = idx; data1_start[rank]++; /* for local proccessing */
179: proc_end = ctable[idx];
180: for (proc_id=0; proc_id<=proc_end; proc_id++){ /* for others to process */
181: if (proc_id == rank ) continue; /* done before this loop */
182: if (proc_id < proc_end && !PetscBTLookup(table[proc_id],idx))
183: continue; /* no need for sending idx to [proc_id] */
184: *data1_start[proc_id] = idx; data1_start[proc_id]++;
185: len_s[proc_id]++;
186: }
187: }
188: /* update header data */
189: for (proc_id=0; proc_id<size; proc_id++){
190: if (proc_id== rank) continue;
191: *(data1 + proc_id*len + 1 + i) = len_s[proc_id] - iwork[proc_id];
192: iwork[proc_id] = len_s[proc_id] ;
193: }
194: ISRestoreIndices(is[i],&idx_i);
195: }
197: nrqs = 0; nrqr = 0;
198: for (i=0; i<size; i++){
199: data1_start[i] = data1 + i*len;
200: if (len_s[i]){
201: nrqs++;
202: len_s[i] += 1 + is_max; /* add no. of header msg */
203: }
204: }
206: for (i=0; i<is_max; i++) {
207: ISDestroy(is[i]);
208: }
209: PetscFree(n);
210: if (ctable){PetscFree(ctable);}
212: /* Determine the number of messages to expect, their lengths, from from-ids */
213: PetscGatherNumberOfMessages(comm,PETSC_NULL,len_s,&nrqr);
214: PetscGatherMessageLengths(comm,nrqs,nrqr,len_s,&id_r1,&len_r1);
215:
216: /* Now post the sends */
217: PetscMalloc(2*size*sizeof(MPI_Request),&s_waits1);
218: s_waits2 = s_waits1 + size;
219: k = 0;
220: for (proc_id=0; proc_id<size; proc_id++){ /* send data1 to processor [proc_id] */
221: if (len_s[proc_id]){
222: MPI_Isend(data1_start[proc_id],len_s[proc_id],MPIU_INT,proc_id,tag1,comm,s_waits1+k);
223: k++;
224: }
225: }
227: /* 2. Receive other's is[] and process. Then send back */
228: /*-----------------------------------------------------*/
229: len = 0;
230: for (i=0; i<nrqr; i++){
231: if (len_r1[i] > len)len = len_r1[i];
232: }
233: PetscFree(len_r1);
234: PetscFree(id_r1);
236: for (proc_id=0; proc_id<size; proc_id++)
237: len_s[proc_id] = iwork[proc_id] = 0;
238:
239: PetscMalloc((len+1)*sizeof(PetscInt),&odata1);
240: PetscMalloc(size*sizeof(PetscInt**),&odata2_ptr);
241: PetscBTCreate(Mbs,otable);
243: len_max = ois_max*(Mbs+1); /* max space storing all is[] for each receive */
244: len_est = 2*len_max; /* estimated space of storing is[] for all receiving messages */
245: PetscMalloc((len_est+1)*sizeof(PetscInt),&odata2);
246: nodata2 = 0; /* nodata2+1: num of PetscMalloc(,&odata2_ptr[]) called */
247: odata2_ptr[nodata2] = odata2;
248: len_unused = len_est; /* unused space in the array odata2_ptr[nodata2]-- needs to be >= len_max */
249:
250: k = 0;
251: while (k < nrqr){
252: /* Receive messages */
253: MPI_Iprobe(MPI_ANY_SOURCE,tag1,comm,&flag,&r_status);
254: if (flag){
255: MPI_Get_count(&r_status,MPIU_INT,&len);
256: proc_id = r_status.MPI_SOURCE;
257: MPI_Irecv(odata1,len,MPIU_INT,proc_id,r_status.MPI_TAG,comm,&r_req);
258: MPI_Wait(&r_req,&r_status);
260: /* Process messages */
261: /* make sure there is enough unused space in odata2 array */
262: if (len_unused < len_max){ /* allocate more space for odata2 */
263: PetscMalloc((len_est+1)*sizeof(PetscInt),&odata2);
264: odata2_ptr[++nodata2] = odata2;
265: len_unused = len_est;
266: }
268: MatIncreaseOverlap_MPISBAIJ_Local(C,odata1,OTHER,odata2,&otable);
269: len = 1 + odata2[0];
270: for (i=0; i<odata2[0]; i++){
271: len += odata2[1 + i];
272: }
274: /* Send messages back */
275: MPI_Isend(odata2,len,MPIU_INT,proc_id,tag2,comm,s_waits2+k);
276: k++;
277: odata2 += len;
278: len_unused -= len;
279: len_s[proc_id] = len; /* num of messages sending back to [proc_id] by this proc */
280: }
281: }
282: PetscFree(odata1);
283: PetscBTDestroy(otable);
285: /* 3. Do local work on this processor's is[] */
286: /*-------------------------------------------*/
287: /* make sure there is enough unused space in odata2(=data) array */
288: len_max = is_max*(Mbs+1); /* max space storing all is[] for this processor */
289: if (len_unused < len_max){ /* allocate more space for odata2 */
290: PetscMalloc((len_est+1)*sizeof(PetscInt),&odata2);
291: odata2_ptr[++nodata2] = odata2;
292: len_unused = len_est;
293: }
295: data = odata2;
296: MatIncreaseOverlap_MPISBAIJ_Local(C,data1_start[rank],MINE,data,table);
297: PetscFree(data1_start);
299: /* 4. Receive work done on other processors, then merge */
300: /*------------------------------------------------------*/
301: /* get max number of messages that this processor expects to recv */
302: MPI_Allreduce(len_s,iwork,size,MPIU_INT,MPI_MAX,comm);
303: PetscMalloc((iwork[rank]+1)*sizeof(PetscInt),&data2);
304: PetscFree(len_s);
306: k = 0;
307: while (k < nrqs){
308: /* Receive messages */
309: MPI_Iprobe(MPI_ANY_SOURCE,tag2,comm,&flag,&r_status);
310: if (flag){
311: MPI_Get_count(&r_status,MPIU_INT,&len);
312: proc_id = r_status.MPI_SOURCE;
313: MPI_Irecv(data2,len,MPIU_INT,proc_id,r_status.MPI_TAG,comm,&r_req);
314: MPI_Wait(&r_req,&r_status);
315: if (len > 1+is_max){ /* Add data2 into data */
316: data2_i = data2 + 1 + is_max;
317: for (i=0; i<is_max; i++){
318: table_i = table[i];
319: data_i = data + 1 + is_max + Mbs*i;
320: isz = data[1+i];
321: for (j=0; j<data2[1+i]; j++){
322: col = data2_i[j];
323: if (!PetscBTLookupSet(table_i,col)) {data_i[isz++] = col;}
324: }
325: data[1+i] = isz;
326: if (i < is_max - 1) data2_i += data2[1+i];
327: }
328: }
329: k++;
330: }
331: }
332: PetscFree(data2);
333: PetscFree(table);
335: /* phase 1 sends are complete */
336: PetscMalloc(size*sizeof(MPI_Status),&s_status);
337: if (nrqs) {MPI_Waitall(nrqs,s_waits1,s_status);}
338: PetscFree(data1);
339:
340: /* phase 2 sends are complete */
341: if (nrqr){MPI_Waitall(nrqr,s_waits2,s_status);}
342: PetscFree(s_waits1);
343: PetscFree(s_status);
345: /* 5. Create new is[] */
346: /*--------------------*/
347: for (i=0; i<is_max; i++) {
348: data_i = data + 1 + is_max + Mbs*i;
349: ISCreateGeneral(PETSC_COMM_SELF,data[1+i],data_i,is+i);
350: }
351: for (k=0; k<=nodata2; k++){
352: PetscFree(odata2_ptr[k]);
353: }
354: PetscFree(odata2_ptr);
356: return(0);
357: }
361: /*
362: MatIncreaseOverlap_MPISBAIJ_Local - Called by MatIncreaseOverlap, to do
363: the work on the local processor.
365: Inputs:
366: C - MAT_MPISBAIJ;
367: data - holds is[]. See MatIncreaseOverlap_MPISBAIJ_Once() for the format.
368: whose - whose is[] to be processed,
369: MINE: this processor's is[]
370: OTHER: other processor's is[]
371: Output:
372: nidx - whose = MINE:
373: holds input and newly found indices in the same format as data
374: whose = OTHER:
375: only holds the newly found indices
376: table - table[i]: mark the indices of is[i], i=0,...,is_max. Used only in the case 'whose=MINE'.
377: */
378: /* Would computation be reduced by swapping the loop 'for each is' and 'for each row'? */
379: static PetscErrorCode MatIncreaseOverlap_MPISBAIJ_Local(Mat C,PetscInt *data,PetscInt whose,PetscInt *nidx,PetscBT *table)
380: {
381: Mat_MPISBAIJ *c = (Mat_MPISBAIJ*)C->data;
382: Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ*)(c->A)->data;
383: Mat_SeqBAIJ *b = (Mat_SeqBAIJ*)(c->B)->data;
385: PetscInt row,mbs,Mbs,*nidx_i,col,col_max,isz,isz0,*ai,*aj,*bi,*bj,*garray,rstart,l;
386: PetscInt a_start,a_end,b_start,b_end,i,j,k,is_max,*idx_i,n;
387: PetscBT table0; /* mark the indices of input is[] for look up */
388: PetscBT table_i; /* poits to i-th table. When whose=OTHER, a single table is used for all is[] */
389:
391: Mbs = c->Mbs; mbs = a->mbs;
392: ai = a->i; aj = a->j;
393: bi = b->i; bj = b->j;
394: garray = c->garray;
395: rstart = c->rstart;
396: is_max = data[0];
398: PetscBTCreate(Mbs,table0);
399:
400: nidx[0] = is_max;
401: idx_i = data + is_max + 1; /* ptr to input is[0] array */
402: nidx_i = nidx + is_max + 1; /* ptr to output is[0] array */
403: for (i=0; i<is_max; i++) { /* for each is */
404: isz = 0;
405: n = data[1+i]; /* size of input is[i] */
407: /* initialize and set table_i(mark idx and nidx) and table0(only mark idx) */
408: if (whose == MINE){ /* process this processor's is[] */
409: table_i = table[i];
410: nidx_i = nidx + 1+ is_max + Mbs*i;
411: } else { /* process other processor's is[] - only use one temp table */
412: table_i = table[0];
413: }
414: PetscBTMemzero(Mbs,table_i);
415: PetscBTMemzero(Mbs,table0);
416: if (n==0) {
417: nidx[1+i] = 0; /* size of new is[i] */
418: continue;
419: }
421: isz0 = 0; col_max = 0;
422: for (j=0; j<n; j++){
423: col = idx_i[j];
424: if (col >= Mbs) SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"index col %D >= Mbs %D",col,Mbs);
425: if(!PetscBTLookupSet(table_i,col)) {
426: PetscBTSet(table0,col);
427: if (whose == MINE) {nidx_i[isz0] = col;}
428: if (col_max < col) col_max = col;
429: isz0++;
430: }
431: }
432:
433: if (whose == MINE) {isz = isz0;}
434: k = 0; /* no. of indices from input is[i] that have been examined */
435: for (row=0; row<mbs; row++){
436: a_start = ai[row]; a_end = ai[row+1];
437: b_start = bi[row]; b_end = bi[row+1];
438: if (PetscBTLookup(table0,row+rstart)){ /* row is on input is[i]:
439: do row search: collect all col in this row */
440: for (l = a_start; l<a_end ; l++){ /* Amat */
441: col = aj[l] + rstart;
442: if (!PetscBTLookupSet(table_i,col)) {nidx_i[isz++] = col;}
443: }
444: for (l = b_start; l<b_end ; l++){ /* Bmat */
445: col = garray[bj[l]];
446: if (!PetscBTLookupSet(table_i,col)) {nidx_i[isz++] = col;}
447: }
448: k++;
449: if (k >= isz0) break; /* for (row=0; row<mbs; row++) */
450: } else { /* row is not on input is[i]:
451: do col serach: add row onto nidx_i if there is a col in nidx_i */
452: for (l = a_start; l<a_end ; l++){ /* Amat */
453: col = aj[l] + rstart;
454: if (col > col_max) break;
455: if (PetscBTLookup(table0,col)){
456: if (!PetscBTLookupSet(table_i,row+rstart)) {nidx_i[isz++] = row+rstart;}
457: break; /* for l = start; l<end ; l++) */
458: }
459: }
460: for (l = b_start; l<b_end ; l++){ /* Bmat */
461: col = garray[bj[l]];
462: if (col > col_max) break;
463: if (PetscBTLookup(table0,col)){
464: if (!PetscBTLookupSet(table_i,row+rstart)) {nidx_i[isz++] = row+rstart;}
465: break; /* for l = start; l<end ; l++) */
466: }
467: }
468: }
469: }
470:
471: if (i < is_max - 1){
472: idx_i += n; /* ptr to input is[i+1] array */
473: nidx_i += isz; /* ptr to output is[i+1] array */
474: }
475: nidx[1+i] = isz; /* size of new is[i] */
476: } /* for each is */
477: PetscBTDestroy(table0);
478:
479: return(0);
480: }