Actual source code: sbaijov.c
petsc-3.9.4 2018-09-11
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*);
12: PetscErrorCode MatIncreaseOverlap_MPISBAIJ(Mat C,PetscInt is_max,IS is[],PetscInt ov)
13: {
15: PetscInt i,N=C->cmap->N, bs=C->rmap->bs,M=C->rmap->N,Mbs=M/bs,*nidx,isz,iov;
16: IS *is_new,*is_row;
17: Mat *submats;
18: Mat_MPISBAIJ *c=(Mat_MPISBAIJ*)C->data;
19: Mat_SeqSBAIJ *asub_i;
20: PetscBT table;
21: PetscInt *ai,brow,nz,nis,l,nmax,nstages_local,nstages,max_no,pos;
22: const PetscInt *idx;
23: PetscBool flg;
26: PetscMalloc1(is_max,&is_new);
27: /* Convert the indices into block format */
28: ISCompressIndicesGeneral(N,C->rmap->n,bs,is_max,is,is_new);
29: if (ov < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative overlap specified\n");
31: /* ----- previous non-scalable implementation ----- */
32: flg = PETSC_FALSE;
33: PetscOptionsHasName(NULL,NULL, "-IncreaseOverlap_old", &flg);
34: if (flg) { /* previous non-scalable implementation */
35: printf("use previous non-scalable implementation...\n");
36: for (i=0; i<ov; ++i) {
37: MatIncreaseOverlap_MPISBAIJ_Once(C,is_max,is_new);
38: }
39: } else { /* implementation using modified BAIJ routines */
41: PetscMalloc1(Mbs+1,&nidx);
42: PetscBTCreate(Mbs,&table); /* for column search */
44: /* Create is_row */
45: PetscMalloc1(is_max,&is_row);
46: ISCreateStride(PETSC_COMM_SELF,Mbs,0,1,&is_row[0]);
48: for (i=1; i<is_max; i++) {
49: is_row[i] = is_row[0]; /* reuse is_row[0] */
50: }
52: /* Allocate memory to hold all the submatrices - Modified from MatCreateSubMatrices_MPIBAIJ() */
53: PetscMalloc1(is_max+1,&submats);
55: /* Determine the number of stages through which submatrices are done */
56: nmax = 20*1000000 / (c->Nbs * sizeof(PetscInt));
57: if (!nmax) nmax = 1;
58: nstages_local = is_max/nmax + ((is_max % nmax) ? 1 : 0);
60: /* Make sure every processor loops through the nstages */
61: MPIU_Allreduce(&nstages_local,&nstages,1,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)C));
63: for (iov=0; iov<ov; ++iov) {
64: /* 1) Get submats for column search */
65: for (i=0,pos=0; i<nstages; i++) {
66: if (pos+nmax <= is_max) max_no = nmax;
67: else if (pos == is_max) max_no = 0;
68: else max_no = is_max-pos;
69: c->ijonly = PETSC_TRUE; /* only matrix data structures are requested */
70: MatCreateSubMatrices_MPIBAIJ_local(C,max_no,is_row+pos,is_new+pos,MAT_INITIAL_MATRIX,submats+pos);
71: pos += max_no;
72: }
74: /* 2) Row search */
75: MatIncreaseOverlap_MPIBAIJ_Once(C,is_max,is_new);
77: /* 3) Column search */
78: for (i=0; i<is_max; i++) {
79: asub_i = (Mat_SeqSBAIJ*)submats[i]->data;
80: ai = asub_i->i;;
82: /* put is_new obtained from MatIncreaseOverlap_MPIBAIJ() to table */
83: PetscBTMemzero(Mbs,table);
85: ISGetIndices(is_new[i],&idx);
86: ISGetLocalSize(is_new[i],&nis);
87: for (l=0; l<nis; l++) {
88: PetscBTSet(table,idx[l]);
89: nidx[l] = idx[l];
90: }
91: isz = nis;
93: /* add column entries to table */
94: for (brow=0; brow<Mbs; brow++) {
95: nz = ai[brow+1] - ai[brow];
96: if (nz) {
97: if (!PetscBTLookupSet(table,brow)) nidx[isz++] = brow;
98: }
99: }
100: ISRestoreIndices(is_new[i],&idx);
101: ISDestroy(&is_new[i]);
103: /* create updated is_new */
104: ISCreateGeneral(PETSC_COMM_SELF,isz,nidx,PETSC_COPY_VALUES,is_new+i);
105: }
107: /* Free tmp spaces */
108: for (i=0; i<is_max; i++) {
109: MatDestroy(&submats[i]);
110: }
111: }
113: PetscBTDestroy(&table);
114: PetscFree(submats);
115: ISDestroy(&is_row[0]);
116: PetscFree(is_row);
117: PetscFree(nidx);
119: }
121: for (i=0; i<is_max; i++) {ISDestroy(&is[i]);}
122: ISExpandIndicesGeneral(N,N,bs,is_max,is_new,is);
124: for (i=0; i<is_max; i++) {ISDestroy(&is_new[i]);}
125: PetscFree(is_new);
126: return(0);
127: }
129: typedef enum {MINE,OTHER} WhoseOwner;
130: /* data1, odata1 and odata2 are packed in the format (for communication):
131: data[0] = is_max, no of is
132: data[1] = size of is[0]
133: ...
134: data[is_max] = size of is[is_max-1]
135: data[is_max + 1] = data(is[0])
136: ...
137: data[is_max+1+sum(size of is[k]), k=0,...,i-1] = data(is[i])
138: ...
139: data2 is packed in the format (for creating output is[]):
140: data[0] = is_max, no of is
141: data[1] = size of is[0]
142: ...
143: data[is_max] = size of is[is_max-1]
144: data[is_max + 1] = data(is[0])
145: ...
146: data[is_max + 1 + Mbs*i) = data(is[i])
147: ...
148: */
149: static PetscErrorCode MatIncreaseOverlap_MPISBAIJ_Once(Mat C,PetscInt is_max,IS is[])
150: {
151: Mat_MPISBAIJ *c = (Mat_MPISBAIJ*)C->data;
153: PetscMPIInt size,rank,tag1,tag2,*len_s,nrqr,nrqs,*id_r1,*len_r1,flag,len,*iwork;
154: const PetscInt *idx_i;
155: PetscInt idx,isz,col,*n,*data1,**data1_start,*data2,*data2_i,*data,*data_i;
156: PetscInt Mbs,i,j,k,*odata1,*odata2;
157: PetscInt proc_id,**odata2_ptr,*ctable=0,*btable,len_max,len_est;
158: PetscInt proc_end=0,len_unused,nodata2;
159: PetscInt ois_max; /* max no of is[] in each of processor */
160: char *t_p;
161: MPI_Comm comm;
162: MPI_Request *s_waits1,*s_waits2,r_req;
163: MPI_Status *s_status,r_status;
164: PetscBT *table; /* mark indices of this processor's is[] */
165: PetscBT table_i;
166: PetscBT otable; /* mark indices of other processors' is[] */
167: PetscInt bs=C->rmap->bs,Bn = c->B->cmap->n,Bnbs = Bn/bs,*Bowners;
168: IS garray_local,garray_gl;
171: PetscObjectGetComm((PetscObject)C,&comm);
172: size = c->size;
173: rank = c->rank;
174: Mbs = c->Mbs;
176: PetscObjectGetNewTag((PetscObject)C,&tag1);
177: PetscObjectGetNewTag((PetscObject)C,&tag2);
179: /* create tables used in
180: step 1: table[i] - mark c->garray of proc [i]
181: step 3: table[i] - mark indices of is[i] when whose=MINE
182: table[0] - mark incideces of is[] when whose=OTHER */
183: len = PetscMax(is_max, size);
184: PetscMalloc2(len,&table,(Mbs/PETSC_BITS_PER_BYTE+1)*len,&t_p);
185: for (i=0; i<len; i++) {
186: table[i] = t_p + (Mbs/PETSC_BITS_PER_BYTE+1)*i;
187: }
189: MPIU_Allreduce(&is_max,&ois_max,1,MPIU_INT,MPI_MAX,comm);
191: /* 1. Send this processor's is[] to other processors */
192: /*---------------------------------------------------*/
193: /* allocate spaces */
194: PetscMalloc1(is_max,&n);
195: len = 0;
196: for (i=0; i<is_max; i++) {
197: ISGetLocalSize(is[i],&n[i]);
198: len += n[i];
199: }
200: if (!len) {
201: is_max = 0;
202: } else {
203: len += 1 + is_max; /* max length of data1 for one processor */
204: }
207: PetscMalloc1(size*len+1,&data1);
208: PetscMalloc1(size,&data1_start);
209: for (i=0; i<size; i++) data1_start[i] = data1 + i*len;
211: PetscMalloc4(size,&len_s,size,&btable,size,&iwork,size+1,&Bowners);
213: /* gather c->garray from all processors */
214: ISCreateGeneral(comm,Bnbs,c->garray,PETSC_COPY_VALUES,&garray_local);
215: ISAllGather(garray_local, &garray_gl);
216: ISDestroy(&garray_local);
217: MPI_Allgather(&Bnbs,1,MPIU_INT,Bowners+1,1,MPIU_INT,comm);
219: Bowners[0] = 0;
220: for (i=0; i<size; i++) Bowners[i+1] += Bowners[i];
222: if (is_max) {
223: /* hash table ctable which maps c->row to proc_id) */
224: PetscMalloc1(Mbs,&ctable);
225: for (proc_id=0,j=0; proc_id<size; proc_id++) {
226: for (; j<C->rmap->range[proc_id+1]/bs; j++) ctable[j] = proc_id;
227: }
229: /* hash tables marking c->garray */
230: ISGetIndices(garray_gl,&idx_i);
231: for (i=0; i<size; i++) {
232: table_i = table[i];
233: PetscBTMemzero(Mbs,table_i);
234: for (j = Bowners[i]; j<Bowners[i+1]; j++) { /* go through B cols of proc[i]*/
235: PetscBTSet(table_i,idx_i[j]);
236: }
237: }
238: ISRestoreIndices(garray_gl,&idx_i);
239: } /* if (is_max) */
240: ISDestroy(&garray_gl);
242: /* evaluate communication - mesg to who, length, and buffer space */
243: for (i=0; i<size; i++) len_s[i] = 0;
245: /* header of data1 */
246: for (proc_id=0; proc_id<size; proc_id++) {
247: iwork[proc_id] = 0;
248: *data1_start[proc_id] = is_max;
249: data1_start[proc_id]++;
250: for (j=0; j<is_max; j++) {
251: if (proc_id == rank) {
252: *data1_start[proc_id] = n[j];
253: } else {
254: *data1_start[proc_id] = 0;
255: }
256: data1_start[proc_id]++;
257: }
258: }
260: for (i=0; i<is_max; i++) {
261: ISGetIndices(is[i],&idx_i);
262: for (j=0; j<n[i]; j++) {
263: idx = idx_i[j];
264: *data1_start[rank] = idx; data1_start[rank]++; /* for local proccessing */
265: proc_end = ctable[idx];
266: for (proc_id=0; proc_id<=proc_end; proc_id++) { /* for others to process */
267: if (proc_id == rank) continue; /* done before this loop */
268: if (proc_id < proc_end && !PetscBTLookup(table[proc_id],idx)) continue; /* no need for sending idx to [proc_id] */
269: *data1_start[proc_id] = idx; data1_start[proc_id]++;
270: len_s[proc_id]++;
271: }
272: }
273: /* update header data */
274: for (proc_id=0; proc_id<size; proc_id++) {
275: if (proc_id== rank) continue;
276: *(data1 + proc_id*len + 1 + i) = len_s[proc_id] - iwork[proc_id];
277: iwork[proc_id] = len_s[proc_id];
278: }
279: ISRestoreIndices(is[i],&idx_i);
280: }
282: nrqs = 0; nrqr = 0;
283: for (i=0; i<size; i++) {
284: data1_start[i] = data1 + i*len;
285: if (len_s[i]) {
286: nrqs++;
287: len_s[i] += 1 + is_max; /* add no. of header msg */
288: }
289: }
291: for (i=0; i<is_max; i++) {
292: ISDestroy(&is[i]);
293: }
294: PetscFree(n);
295: PetscFree(ctable);
297: /* Determine the number of messages to expect, their lengths, from from-ids */
298: PetscGatherNumberOfMessages(comm,NULL,len_s,&nrqr);
299: PetscGatherMessageLengths(comm,nrqs,nrqr,len_s,&id_r1,&len_r1);
301: /* Now post the sends */
302: PetscMalloc2(size,&s_waits1,size,&s_waits2);
303: k = 0;
304: for (proc_id=0; proc_id<size; proc_id++) { /* send data1 to processor [proc_id] */
305: if (len_s[proc_id]) {
306: MPI_Isend(data1_start[proc_id],len_s[proc_id],MPIU_INT,proc_id,tag1,comm,s_waits1+k);
307: k++;
308: }
309: }
311: /* 2. Receive other's is[] and process. Then send back */
312: /*-----------------------------------------------------*/
313: len = 0;
314: for (i=0; i<nrqr; i++) {
315: if (len_r1[i] > len) len = len_r1[i];
316: }
317: PetscFree(len_r1);
318: PetscFree(id_r1);
320: for (proc_id=0; proc_id<size; proc_id++) len_s[proc_id] = iwork[proc_id] = 0;
322: PetscMalloc1(len+1,&odata1);
323: PetscMalloc1(size,&odata2_ptr);
324: PetscBTCreate(Mbs,&otable);
326: len_max = ois_max*(Mbs+1); /* max space storing all is[] for each receive */
327: len_est = 2*len_max; /* estimated space of storing is[] for all receiving messages */
328: PetscMalloc1(len_est+1,&odata2);
329: nodata2 = 0; /* nodata2+1: num of PetscMalloc(,&odata2_ptr[]) called */
331: odata2_ptr[nodata2] = odata2;
333: len_unused = len_est; /* unused space in the array odata2_ptr[nodata2]-- needs to be >= len_max */
335: k = 0;
336: while (k < nrqr) {
337: /* Receive messages */
338: MPI_Iprobe(MPI_ANY_SOURCE,tag1,comm,&flag,&r_status);
339: if (flag) {
340: MPI_Get_count(&r_status,MPIU_INT,&len);
341: proc_id = r_status.MPI_SOURCE;
342: MPI_Irecv(odata1,len,MPIU_INT,proc_id,r_status.MPI_TAG,comm,&r_req);
343: MPI_Wait(&r_req,&r_status);
345: /* Process messages */
346: /* make sure there is enough unused space in odata2 array */
347: if (len_unused < len_max) { /* allocate more space for odata2 */
348: PetscMalloc1(len_est+1,&odata2);
350: odata2_ptr[++nodata2] = odata2;
352: len_unused = len_est;
353: }
355: MatIncreaseOverlap_MPISBAIJ_Local(C,odata1,OTHER,odata2,&otable);
356: len = 1 + odata2[0];
357: for (i=0; i<odata2[0]; i++) len += odata2[1 + i];
359: /* Send messages back */
360: MPI_Isend(odata2,len,MPIU_INT,proc_id,tag2,comm,s_waits2+k);
361: k++;
362: odata2 += len;
363: len_unused -= len;
364: len_s[proc_id] = len; /* num of messages sending back to [proc_id] by this proc */
365: }
366: }
367: PetscFree(odata1);
368: PetscBTDestroy(&otable);
370: /* 3. Do local work on this processor's is[] */
371: /*-------------------------------------------*/
372: /* make sure there is enough unused space in odata2(=data) array */
373: len_max = is_max*(Mbs+1); /* max space storing all is[] for this processor */
374: if (len_unused < len_max) { /* allocate more space for odata2 */
375: PetscMalloc1(len_est+1,&odata2);
377: odata2_ptr[++nodata2] = odata2;
378: }
380: data = odata2;
381: MatIncreaseOverlap_MPISBAIJ_Local(C,data1_start[rank],MINE,data,table);
382: PetscFree(data1_start);
384: /* 4. Receive work done on other processors, then merge */
385: /*------------------------------------------------------*/
386: /* get max number of messages that this processor expects to recv */
387: MPIU_Allreduce(len_s,iwork,size,MPI_INT,MPI_MAX,comm);
388: PetscMalloc1(iwork[rank]+1,&data2);
389: PetscFree4(len_s,btable,iwork,Bowners);
391: k = 0;
392: while (k < nrqs) {
393: /* Receive messages */
394: MPI_Iprobe(MPI_ANY_SOURCE,tag2,comm,&flag,&r_status);
395: if (flag) {
396: MPI_Get_count(&r_status,MPIU_INT,&len);
398: proc_id = r_status.MPI_SOURCE;
400: MPI_Irecv(data2,len,MPIU_INT,proc_id,r_status.MPI_TAG,comm,&r_req);
401: MPI_Wait(&r_req,&r_status);
402: if (len > 1+is_max) { /* Add data2 into data */
403: data2_i = data2 + 1 + is_max;
404: for (i=0; i<is_max; i++) {
405: table_i = table[i];
406: data_i = data + 1 + is_max + Mbs*i;
407: isz = data[1+i];
408: for (j=0; j<data2[1+i]; j++) {
409: col = data2_i[j];
410: if (!PetscBTLookupSet(table_i,col)) data_i[isz++] = col;
411: }
412: data[1+i] = isz;
413: if (i < is_max - 1) data2_i += data2[1+i];
414: }
415: }
416: k++;
417: }
418: }
419: PetscFree(data2);
420: PetscFree2(table,t_p);
422: /* phase 1 sends are complete */
423: PetscMalloc1(size,&s_status);
424: if (nrqs) {MPI_Waitall(nrqs,s_waits1,s_status);}
425: PetscFree(data1);
427: /* phase 2 sends are complete */
428: if (nrqr) {MPI_Waitall(nrqr,s_waits2,s_status);}
429: PetscFree2(s_waits1,s_waits2);
430: PetscFree(s_status);
432: /* 5. Create new is[] */
433: /*--------------------*/
434: for (i=0; i<is_max; i++) {
435: data_i = data + 1 + is_max + Mbs*i;
436: ISCreateGeneral(PETSC_COMM_SELF,data[1+i],data_i,PETSC_COPY_VALUES,is+i);
437: }
438: for (k=0; k<=nodata2; k++) {
439: PetscFree(odata2_ptr[k]);
440: }
441: PetscFree(odata2_ptr);
442: return(0);
443: }
445: /*
446: MatIncreaseOverlap_MPISBAIJ_Local - Called by MatIncreaseOverlap, to do
447: the work on the local processor.
449: Inputs:
450: C - MAT_MPISBAIJ;
451: data - holds is[]. See MatIncreaseOverlap_MPISBAIJ_Once() for the format.
452: whose - whose is[] to be processed,
453: MINE: this processor's is[]
454: OTHER: other processor's is[]
455: Output:
456: nidx - whose = MINE:
457: holds input and newly found indices in the same format as data
458: whose = OTHER:
459: only holds the newly found indices
460: table - table[i]: mark the indices of is[i], i=0,...,is_max. Used only in the case 'whose=MINE'.
461: */
462: /* Would computation be reduced by swapping the loop 'for each is' and 'for each row'? */
463: static PetscErrorCode MatIncreaseOverlap_MPISBAIJ_Local(Mat C,PetscInt *data,PetscInt whose,PetscInt *nidx,PetscBT *table)
464: {
465: Mat_MPISBAIJ *c = (Mat_MPISBAIJ*)C->data;
466: Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ*)(c->A)->data;
467: Mat_SeqBAIJ *b = (Mat_SeqBAIJ*)(c->B)->data;
469: PetscInt row,mbs,Mbs,*nidx_i,col,col_max,isz,isz0,*ai,*aj,*bi,*bj,*garray,rstart,l;
470: PetscInt a_start,a_end,b_start,b_end,i,j,k,is_max,*idx_i,n;
471: PetscBT table0; /* mark the indices of input is[] for look up */
472: PetscBT table_i; /* poits to i-th table. When whose=OTHER, a single table is used for all is[] */
475: Mbs = c->Mbs; mbs = a->mbs;
476: ai = a->i; aj = a->j;
477: bi = b->i; bj = b->j;
478: garray = c->garray;
479: rstart = c->rstartbs;
480: is_max = data[0];
482: PetscBTCreate(Mbs,&table0);
484: nidx[0] = is_max;
485: idx_i = data + is_max + 1; /* ptr to input is[0] array */
486: nidx_i = nidx + is_max + 1; /* ptr to output is[0] array */
487: for (i=0; i<is_max; i++) { /* for each is */
488: isz = 0;
489: n = data[1+i]; /* size of input is[i] */
491: /* initialize and set table_i(mark idx and nidx) and table0(only mark idx) */
492: if (whose == MINE) { /* process this processor's is[] */
493: table_i = table[i];
494: nidx_i = nidx + 1+ is_max + Mbs*i;
495: } else { /* process other processor's is[] - only use one temp table */
496: table_i = table[0];
497: }
498: PetscBTMemzero(Mbs,table_i);
499: PetscBTMemzero(Mbs,table0);
500: if (n==0) {
501: nidx[1+i] = 0; /* size of new is[i] */
502: continue;
503: }
505: isz0 = 0; col_max = 0;
506: for (j=0; j<n; j++) {
507: col = idx_i[j];
508: if (col >= Mbs) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"index col %D >= Mbs %D",col,Mbs);
509: if (!PetscBTLookupSet(table_i,col)) {
510: PetscBTSet(table0,col);
511: if (whose == MINE) nidx_i[isz0] = col;
512: if (col_max < col) col_max = col;
513: isz0++;
514: }
515: }
517: if (whose == MINE) isz = isz0;
518: k = 0; /* no. of indices from input is[i] that have been examined */
519: for (row=0; row<mbs; row++) {
520: a_start = ai[row]; a_end = ai[row+1];
521: b_start = bi[row]; b_end = bi[row+1];
522: if (PetscBTLookup(table0,row+rstart)) { /* row is on input is[i]:
523: do row search: collect all col in this row */
524: for (l = a_start; l<a_end ; l++) { /* Amat */
525: col = aj[l] + rstart;
526: if (!PetscBTLookupSet(table_i,col)) nidx_i[isz++] = col;
527: }
528: for (l = b_start; l<b_end ; l++) { /* Bmat */
529: col = garray[bj[l]];
530: if (!PetscBTLookupSet(table_i,col)) nidx_i[isz++] = col;
531: }
532: k++;
533: if (k >= isz0) break; /* for (row=0; row<mbs; row++) */
534: } else { /* row is not on input is[i]:
535: do col serach: add row onto nidx_i if there is a col in nidx_i */
536: for (l = a_start; l<a_end; l++) { /* Amat */
537: col = aj[l] + rstart;
538: if (col > col_max) break;
539: if (PetscBTLookup(table0,col)) {
540: if (!PetscBTLookupSet(table_i,row+rstart)) nidx_i[isz++] = row+rstart;
541: break; /* for l = start; l<end ; l++) */
542: }
543: }
544: for (l = b_start; l<b_end; l++) { /* Bmat */
545: col = garray[bj[l]];
546: if (col > col_max) break;
547: if (PetscBTLookup(table0,col)) {
548: if (!PetscBTLookupSet(table_i,row+rstart)) nidx_i[isz++] = row+rstart;
549: break; /* for l = start; l<end ; l++) */
550: }
551: }
552: }
553: }
555: if (i < is_max - 1) {
556: idx_i += n; /* ptr to input is[i+1] array */
557: nidx_i += isz; /* ptr to output is[i+1] array */
558: }
559: nidx[1+i] = isz; /* size of new is[i] */
560: } /* for each is */
561: PetscBTDestroy(&table0);
562: return(0);
563: }