Actual source code: sbaijov.c
2: /*
3: Routines to compute overlapping regions of a parallel MPI matrix.
4: Used for finding submatrices that were shared across processors.
5: */
6: #include src/mat/impls/sbaij/mpi/mpisbaij.h
7: #include petscbt.h
9: static PetscErrorCode MatIncreaseOverlap_MPISBAIJ_Once(Mat,PetscInt,IS*);
10: static PetscErrorCode MatIncreaseOverlap_MPISBAIJ_Local(Mat,PetscInt*,PetscInt,PetscInt*,PetscBT*);
14: PetscErrorCode MatIncreaseOverlap_MPISBAIJ(Mat C,PetscInt is_max,IS is[],PetscInt ov)
15: {
17: PetscInt i,N=C->N, bs=C->bs;
18: IS *is_new;
21: PetscMalloc(is_max*sizeof(IS),&is_new);
22: /* Convert the indices into block format */
23: ISCompressIndicesGeneral(N,bs,is_max,is,is_new);
24: if (ov < 0){ SETERRQ(PETSC_ERR_ARG_OUTOFRANGE,"Negative overlap specified\n");}
25: for (i=0; i<ov; ++i) {
26: MatIncreaseOverlap_MPISBAIJ_Once(C,is_max,is_new);
27: }
28: for (i=0; i<is_max; i++) {ISDestroy(is[i]);}
29: ISExpandIndicesGeneral(N,bs,is_max,is_new,is);
30: for (i=0; i<is_max; i++) {ISDestroy(is_new[i]);}
31: PetscFree(is_new);
32: return(0);
33: }
35: typedef enum {MINE,OTHER} WhoseOwner;
36: /* data1, odata1 and odata2 are packed in the format (for communication):
37: data[0] = is_max, no of is
38: data[1] = size of is[0]
39: ...
40: data[is_max] = size of is[is_max-1]
41: data[is_max + 1] = data(is[0])
42: ...
43: data[is_max+1+sum(size of is[k]), k=0,...,i-1] = data(is[i])
44: ...
45: data2 is packed in the format (for creating output is[]):
46: data[0] = is_max, no of is
47: data[1] = size of is[0]
48: ...
49: data[is_max] = size of is[is_max-1]
50: data[is_max + 1] = data(is[0])
51: ...
52: data[is_max + 1 + Mbs*i) = data(is[i])
53: ...
54: */
57: static PetscErrorCode MatIncreaseOverlap_MPISBAIJ_Once(Mat C,PetscInt is_max,IS is[])
58: {
59: Mat_MPISBAIJ *c = (Mat_MPISBAIJ*)C->data;
61: PetscMPIInt size,rank,tag1,tag2,*len_s,nrqr,nrqs,*id_r1,*len_r1,flag,len;
62: PetscInt idx,*idx_i,isz,col,*n,*data1,**data1_start,*data2,*data2_i,*data,*data_i,
63: Mbs,i,j,k,*odata1,*odata2,
64: proc_id,**odata2_ptr,*ctable=0,*btable,len_max,len_est;
65: PetscInt proc_end=0,*iwork,len_unused,nodata2;
66: PetscInt ois_max; /* max no of is[] in each of processor */
67: char *t_p;
68: MPI_Comm comm;
69: MPI_Request *s_waits1,*s_waits2,r_req;
70: MPI_Status *s_status,r_status;
71: PetscBT *table; /* mark indices of this processor's is[] */
72: PetscBT table_i;
73: PetscBT otable; /* mark indices of other processors' is[] */
74: PetscInt bs=C->bs,Bn = c->B->n,Bnbs = Bn/bs,*Bowners;
75: IS garray_local,garray_gl;
78: comm = C->comm;
79: size = c->size;
80: rank = c->rank;
81: Mbs = c->Mbs;
83: PetscObjectGetNewTag((PetscObject)C,&tag1);
84: PetscObjectGetNewTag((PetscObject)C,&tag2);
86: /* create tables used in
87: step 1: table[i] - mark c->garray of proc [i]
88: step 3: table[i] - mark indices of is[i] when whose=MINE
89: table[0] - mark incideces of is[] when whose=OTHER */
90: len = PetscMax(is_max, size);
91: len_max = len*sizeof(PetscBT) + (Mbs/PETSC_BITS_PER_BYTE+1)*len*sizeof(char) + 1;
92: PetscMalloc(len_max,&table);
93: t_p = (char *)(table + len);
94: for (i=0; i<len; i++) {
95: table[i] = t_p + (Mbs/PETSC_BITS_PER_BYTE+1)*i;
96: }
98: MPI_Allreduce(&is_max,&ois_max,1,MPIU_INT,MPI_MAX,comm);
99:
100: /* 1. Send this processor's is[] to other processors */
101: /*---------------------------------------------------*/
102: /* allocate spaces */
103: PetscMalloc(is_max*sizeof(PetscInt),&n);
104: len = 0;
105: for (i=0; i<is_max; i++) {
106: ISGetLocalSize(is[i],&n[i]);
107: len += n[i];
108: }
109: if (!len) {
110: is_max = 0;
111: } else {
112: len += 1 + is_max; /* max length of data1 for one processor */
113: }
115:
116: PetscMalloc((size*len+1)*sizeof(PetscInt),&data1);
117: PetscMalloc(size*sizeof(PetscInt*),&data1_start);
118: for (i=0; i<size; i++) data1_start[i] = data1 + i*len;
120: PetscMalloc((size*4+1)*sizeof(PetscInt),&len_s);
121: btable = (PetscInt*)(len_s + size);
122: iwork = btable + size;
123: Bowners = iwork + size;
125: /* gather c->garray from all processors */
126: ISCreateGeneral(comm,Bnbs,c->garray,&garray_local);
127: ISAllGather(garray_local, &garray_gl);
128: ISDestroy(garray_local);
129: MPI_Allgather(&Bnbs,1,MPIU_INT,Bowners+1,1,MPIU_INT,comm);
130: Bowners[0] = 0;
131: for (i=0; i<size; i++) Bowners[i+1] += Bowners[i];
132:
133: if (is_max){
134: /* hash table ctable which maps c->row to proc_id) */
135: PetscMalloc(Mbs*sizeof(PetscInt),&ctable);
136: for (proc_id=0,j=0; proc_id<size; proc_id++) {
137: for (; j<c->rowners[proc_id+1]; j++) {
138: ctable[j] = proc_id;
139: }
140: }
142: /* hash tables marking c->garray */
143: ISGetIndices(garray_gl,&idx_i);
144: for (i=0; i<size; i++){
145: table_i = table[i];
146: PetscBTMemzero(Mbs,table_i);
147: for (j = Bowners[i]; j<Bowners[i+1]; j++){ /* go through B cols of proc[i]*/
148: PetscBTSet(table_i,idx_i[j]);
149: }
150: }
151: ISRestoreIndices(garray_gl,&idx_i);
152: } /* if (is_max) */
153: ISDestroy(garray_gl);
155: /* evaluate communication - mesg to who, length, and buffer space */
156: for (i=0; i<size; i++) len_s[i] = 0;
157:
158: /* header of data1 */
159: for (proc_id=0; proc_id<size; proc_id++){
160: iwork[proc_id] = 0;
161: *data1_start[proc_id] = is_max;
162: data1_start[proc_id]++;
163: for (j=0; j<is_max; j++) {
164: if (proc_id == rank){
165: *data1_start[proc_id] = n[j];
166: } else {
167: *data1_start[proc_id] = 0;
168: }
169: data1_start[proc_id]++;
170: }
171: }
172:
173: for (i=0; i<is_max; i++) {
174: ISGetIndices(is[i],&idx_i);
175: for (j=0; j<n[i]; j++){
176: idx = idx_i[j];
177: *data1_start[rank] = idx; data1_start[rank]++; /* for local proccessing */
178: proc_end = ctable[idx];
179: for (proc_id=0; proc_id<=proc_end; proc_id++){ /* for others to process */
180: if (proc_id == rank ) continue; /* done before this loop */
181: if (proc_id < proc_end && !PetscBTLookup(table[proc_id],idx))
182: continue; /* no need for sending idx to [proc_id] */
183: *data1_start[proc_id] = idx; data1_start[proc_id]++;
184: len_s[proc_id]++;
185: }
186: }
187: /* update header data */
188: for (proc_id=0; proc_id<size; proc_id++){
189: if (proc_id== rank) continue;
190: *(data1 + proc_id*len + 1 + i) = len_s[proc_id] - iwork[proc_id];
191: iwork[proc_id] = len_s[proc_id] ;
192: }
193: ISRestoreIndices(is[i],&idx_i);
194: }
196: nrqs = 0; nrqr = 0;
197: for (i=0; i<size; i++){
198: data1_start[i] = data1 + i*len;
199: if (len_s[i]){
200: nrqs++;
201: len_s[i] += 1 + is_max; /* add no. of header msg */
202: }
203: }
205: for (i=0; i<is_max; i++) {
206: ISDestroy(is[i]);
207: }
208: PetscFree(n);
209: if (ctable){PetscFree(ctable);}
211: /* Determine the number of messages to expect, their lengths, from from-ids */
212: PetscGatherNumberOfMessages(comm,PETSC_NULL,len_s,&nrqr);
213: PetscGatherMessageLengths(comm,nrqs,nrqr,len_s,&id_r1,&len_r1);
214:
215: /* Now post the sends */
216: PetscMalloc(2*size*sizeof(MPI_Request),&s_waits1);
217: s_waits2 = s_waits1 + size;
218: k = 0;
219: for (proc_id=0; proc_id<size; proc_id++){ /* send data1 to processor [proc_id] */
220: if (len_s[proc_id]){
221: MPI_Isend(data1_start[proc_id],len_s[proc_id],MPIU_INT,proc_id,tag1,comm,s_waits1+k);
222: k++;
223: }
224: }
226: /* 2. Receive other's is[] and process. Then send back */
227: /*-----------------------------------------------------*/
228: len = 0;
229: for (i=0; i<nrqr; i++){
230: if (len_r1[i] > len)len = len_r1[i];
231: }
232: PetscFree(len_r1);
233: PetscFree(id_r1);
235: for (proc_id=0; proc_id<size; proc_id++)
236: len_s[proc_id] = iwork[proc_id] = 0;
237:
238: PetscMalloc((len+1)*sizeof(PetscInt),&odata1);
239: PetscMalloc(size*sizeof(PetscInt**),&odata2_ptr);
240: PetscBTCreate(Mbs,otable);
242: len_max = ois_max*(Mbs+1); /* max space storing all is[] for each receive */
243: len_est = 2*len_max; /* estimated space of storing is[] for all receiving messages */
244: PetscMalloc((len_est+1)*sizeof(PetscInt),&odata2);
245: nodata2 = 0; /* nodata2+1: num of PetscMalloc(,&odata2_ptr[]) called */
246: odata2_ptr[nodata2] = odata2;
247: len_unused = len_est; /* unused space in the array odata2_ptr[nodata2]-- needs to be >= len_max */
248:
249: k = 0;
250: while (k < nrqr){
251: /* Receive messages */
252: MPI_Iprobe(MPI_ANY_SOURCE,tag1,comm,&flag,&r_status);
253: if (flag){
254: MPI_Get_count(&r_status,MPIU_INT,&len);
255: proc_id = r_status.MPI_SOURCE;
256: MPI_Irecv(odata1,len,MPIU_INT,proc_id,r_status.MPI_TAG,comm,&r_req);
257: MPI_Wait(&r_req,&r_status);
259: /* Process messages */
260: /* make sure there is enough unused space in odata2 array */
261: if (len_unused < len_max){ /* allocate more space for odata2 */
262: PetscMalloc((len_est+1)*sizeof(PetscInt),&odata2);
263: odata2_ptr[++nodata2] = odata2;
264: len_unused = len_est;
265: }
267: MatIncreaseOverlap_MPISBAIJ_Local(C,odata1,OTHER,odata2,&otable);
268: len = 1 + odata2[0];
269: for (i=0; i<odata2[0]; i++){
270: len += odata2[1 + i];
271: }
273: /* Send messages back */
274: MPI_Isend(odata2,len,MPIU_INT,proc_id,tag2,comm,s_waits2+k);
275: k++;
276: odata2 += len;
277: len_unused -= len;
278: len_s[proc_id] = len; /* num of messages sending back to [proc_id] by this proc */
279: }
280: }
281: PetscFree(odata1);
282: PetscBTDestroy(otable);
284: /* 3. Do local work on this processor's is[] */
285: /*-------------------------------------------*/
286: /* make sure there is enough unused space in odata2(=data) array */
287: len_max = is_max*(Mbs+1); /* max space storing all is[] for this processor */
288: if (len_unused < len_max){ /* allocate more space for odata2 */
289: PetscMalloc((len_est+1)*sizeof(PetscInt),&odata2);
290: odata2_ptr[++nodata2] = odata2;
291: len_unused = len_est;
292: }
294: data = odata2;
295: MatIncreaseOverlap_MPISBAIJ_Local(C,data1_start[rank],MINE,data,table);
296: PetscFree(data1_start);
298: /* 4. Receive work done on other processors, then merge */
299: /*------------------------------------------------------*/
300: /* get max number of messages that this processor expects to recv */
301: MPI_Allreduce(len_s,iwork,size,MPIU_INT,MPI_MAX,comm);
302: PetscMalloc((iwork[rank]+1)*sizeof(PetscInt),&data2);
303: PetscFree(len_s);
305: k = 0;
306: while (k < nrqs){
307: /* Receive messages */
308: MPI_Iprobe(MPI_ANY_SOURCE,tag2,comm,&flag,&r_status);
309: if (flag){
310: MPI_Get_count(&r_status,MPIU_INT,&len);
311: proc_id = r_status.MPI_SOURCE;
312: MPI_Irecv(data2,len,MPIU_INT,proc_id,r_status.MPI_TAG,comm,&r_req);
313: MPI_Wait(&r_req,&r_status);
314: if (len > 1+is_max){ /* Add data2 into data */
315: data2_i = data2 + 1 + is_max;
316: for (i=0; i<is_max; i++){
317: table_i = table[i];
318: data_i = data + 1 + is_max + Mbs*i;
319: isz = data[1+i];
320: for (j=0; j<data2[1+i]; j++){
321: col = data2_i[j];
322: if (!PetscBTLookupSet(table_i,col)) {data_i[isz++] = col;}
323: }
324: data[1+i] = isz;
325: if (i < is_max - 1) data2_i += data2[1+i];
326: }
327: }
328: k++;
329: }
330: }
331: PetscFree(data2);
332: PetscFree(table);
334: /* phase 1 sends are complete */
335: PetscMalloc(size*sizeof(MPI_Status),&s_status);
336: if (nrqs){
337: MPI_Waitall(nrqs,s_waits1,s_status);
338: }
339: PetscFree(data1);
340:
341: /* phase 2 sends are complete */
342: if (nrqr){
343: MPI_Waitall(nrqr,s_waits2,s_status);
344: }
345: PetscFree(s_waits1);
346: PetscFree(s_status);
348: /* 5. Create new is[] */
349: /*--------------------*/
350: for (i=0; i<is_max; i++) {
351: data_i = data + 1 + is_max + Mbs*i;
352: ISCreateGeneral(PETSC_COMM_SELF,data[1+i],data_i,is+i);
353: }
354: for (k=0; k<=nodata2; k++){
355: PetscFree(odata2_ptr[k]);
356: }
357: PetscFree(odata2_ptr);
359: return(0);
360: }
364: /*
365: MatIncreaseOverlap_MPISBAIJ_Local - Called by MatIncreaseOverlap, to do
366: the work on the local processor.
368: Inputs:
369: C - MAT_MPISBAIJ;
370: data - holds is[]. See MatIncreaseOverlap_MPISBAIJ_Once() for the format.
371: whose - whose is[] to be processed,
372: MINE: this processor's is[]
373: OTHER: other processor's is[]
374: Output:
375: nidx - whose = MINE:
376: holds input and newly found indices in the same format as data
377: whose = OTHER:
378: only holds the newly found indices
379: table - table[i]: mark the indices of is[i], i=0,...,is_max. Used only in the case 'whose=MINE'.
380: */
381: /* Would computation be reduced by swapping the loop 'for each is' and 'for each row'? */
382: static PetscErrorCode MatIncreaseOverlap_MPISBAIJ_Local(Mat C,PetscInt *data,PetscInt whose,PetscInt *nidx,PetscBT *table)
383: {
384: Mat_MPISBAIJ *c = (Mat_MPISBAIJ*)C->data;
385: Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ*)(c->A)->data;
386: Mat_SeqBAIJ *b = (Mat_SeqBAIJ*)(c->B)->data;
388: PetscInt row,mbs,Mbs,*nidx_i,col,col_max,isz,isz0,*ai,*aj,*bi,*bj,*garray,rstart,l;
389: PetscInt a_start,a_end,b_start,b_end,i,j,k,is_max,*idx_i,n;
390: PetscBT table0; /* mark the indices of input is[] for look up */
391: PetscBT table_i; /* poits to i-th table. When whose=OTHER, a single table is used for all is[] */
392:
394: Mbs = c->Mbs; mbs = a->mbs;
395: ai = a->i; aj = a->j;
396: bi = b->i; bj = b->j;
397: garray = c->garray;
398: rstart = c->rstart;
399: is_max = data[0];
401: PetscBTCreate(Mbs,table0);
402:
403: nidx[0] = is_max;
404: idx_i = data + is_max + 1; /* ptr to input is[0] array */
405: nidx_i = nidx + is_max + 1; /* ptr to output is[0] array */
406: for (i=0; i<is_max; i++) { /* for each is */
407: isz = 0;
408: n = data[1+i]; /* size of input is[i] */
410: /* initialize and set table_i(mark idx and nidx) and table0(only mark idx) */
411: if (whose == MINE){ /* process this processor's is[] */
412: table_i = table[i];
413: nidx_i = nidx + 1+ is_max + Mbs*i;
414: } else { /* process other processor's is[] - only use one temp table */
415: table_i = table[0];
416: }
417: PetscBTMemzero(Mbs,table_i);
418: PetscBTMemzero(Mbs,table0);
419: if (n==0) {
420: nidx[1+i] = 0; /* size of new is[i] */
421: continue;
422: }
424: isz0 = 0; col_max = 0;
425: for (j=0; j<n; j++){
426: col = idx_i[j];
427: if (col >= Mbs) SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"index col %D >= Mbs %D",col,Mbs);
428: if(!PetscBTLookupSet(table_i,col)) {
429: PetscBTSet(table0,col);
430: if (whose == MINE) {nidx_i[isz0] = col;}
431: if (col_max < col) col_max = col;
432: isz0++;
433: }
434: }
435:
436: if (whose == MINE) {isz = isz0;}
437: k = 0; /* no. of indices from input is[i] that have been examined */
438: for (row=0; row<mbs; row++){
439: a_start = ai[row]; a_end = ai[row+1];
440: b_start = bi[row]; b_end = bi[row+1];
441: if (PetscBTLookup(table0,row+rstart)){ /* row is on input is[i]:
442: do row search: collect all col in this row */
443: for (l = a_start; l<a_end ; l++){ /* Amat */
444: col = aj[l] + rstart;
445: if (!PetscBTLookupSet(table_i,col)) {nidx_i[isz++] = col;}
446: }
447: for (l = b_start; l<b_end ; l++){ /* Bmat */
448: col = garray[bj[l]];
449: if (!PetscBTLookupSet(table_i,col)) {nidx_i[isz++] = col;}
450: }
451: k++;
452: if (k >= isz0) break; /* for (row=0; row<mbs; row++) */
453: } else { /* row is not on input is[i]:
454: do col serach: add row onto nidx_i if there is a col in nidx_i */
455: for (l = a_start; l<a_end ; l++){ /* Amat */
456: col = aj[l] + rstart;
457: if (col > col_max) break;
458: if (PetscBTLookup(table0,col)){
459: if (!PetscBTLookupSet(table_i,row+rstart)) {nidx_i[isz++] = row+rstart;}
460: break; /* for l = start; l<end ; l++) */
461: }
462: }
463: for (l = b_start; l<b_end ; l++){ /* Bmat */
464: col = garray[bj[l]];
465: if (col > col_max) break;
466: if (PetscBTLookup(table0,col)){
467: if (!PetscBTLookupSet(table_i,row+rstart)) {nidx_i[isz++] = row+rstart;}
468: break; /* for l = start; l<end ; l++) */
469: }
470: }
471: }
472: }
473:
474: if (i < is_max - 1){
475: idx_i += n; /* ptr to input is[i+1] array */
476: nidx_i += isz; /* ptr to output is[i+1] array */
477: }
478: nidx[1+i] = isz; /* size of new is[i] */
479: } /* for each is */
480: PetscBTDestroy(table0);
481:
482: return(0);
483: }