Actual source code: mmdense.c


  2: /*
  3:    Support for the parallel dense matrix vector multiply
  4: */
  5: #include <../src/mat/impls/dense/mpi/mpidense.h>
  6: #include <petscblaslapack.h>

  8: PetscErrorCode MatSetUpMultiply_MPIDense(Mat mat)
  9: {
 10:   Mat_MPIDense *mdn = (Mat_MPIDense *)mat->data;

 12:   PetscFunctionBegin;
 13:   if (!mdn->Mvctx) {
 14:     /* Create local vector that is used to scatter into */
 15:     PetscCall(VecDestroy(&mdn->lvec));
 16:     if (mdn->A) { PetscCall(MatCreateVecs(mdn->A, &mdn->lvec, NULL)); }
 17:     PetscCall(PetscLayoutSetUp(mat->cmap));
 18:     PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)mat), &mdn->Mvctx));
 19:     PetscCall(PetscSFSetGraphWithPattern(mdn->Mvctx, mat->cmap, PETSCSF_PATTERN_ALLGATHER));
 20:   }
 21:   PetscFunctionReturn(PETSC_SUCCESS);
 22: }

 24: static PetscErrorCode MatCreateSubMatrices_MPIDense_Local(Mat, PetscInt, const IS[], const IS[], MatReuse, Mat *);

 26: PetscErrorCode MatCreateSubMatrices_MPIDense(Mat C, PetscInt ismax, const IS isrow[], const IS iscol[], MatReuse scall, Mat *submat[])
 27: {
 28:   PetscInt nmax, nstages_local, nstages, i, pos, max_no;

 30:   PetscFunctionBegin;
 31:   /* Allocate memory to hold all the submatrices */
 32:   if (scall != MAT_REUSE_MATRIX) PetscCall(PetscCalloc1(ismax + 1, submat));
 33:   /* Determine the number of stages through which submatrices are done */
 34:   nmax = 20 * 1000000 / (C->cmap->N * sizeof(PetscInt));
 35:   if (!nmax) nmax = 1;
 36:   nstages_local = ismax / nmax + ((ismax % nmax) ? 1 : 0);

 38:   /* Make sure every processor loops through the nstages */
 39:   PetscCall(MPIU_Allreduce(&nstages_local, &nstages, 1, MPIU_INT, MPI_MAX, PetscObjectComm((PetscObject)C)));

 41:   for (i = 0, pos = 0; i < nstages; i++) {
 42:     if (pos + nmax <= ismax) max_no = nmax;
 43:     else if (pos == ismax) max_no = 0;
 44:     else max_no = ismax - pos;
 45:     PetscCall(MatCreateSubMatrices_MPIDense_Local(C, max_no, isrow + pos, iscol + pos, scall, *submat + pos));
 46:     pos += max_no;
 47:   }
 48:   PetscFunctionReturn(PETSC_SUCCESS);
 49: }

 51: PetscErrorCode MatCreateSubMatrices_MPIDense_Local(Mat C, PetscInt ismax, const IS isrow[], const IS iscol[], MatReuse scall, Mat *submats)
 52: {
 53:   Mat_MPIDense    *c = (Mat_MPIDense *)C->data;
 54:   Mat              A = c->A;
 55:   Mat_SeqDense    *a = (Mat_SeqDense *)A->data, *mat;
 56:   PetscMPIInt      rank, size, tag0, tag1, idex, end, i;
 57:   PetscInt         N = C->cmap->N, rstart = C->rmap->rstart, count;
 58:   const PetscInt **irow, **icol, *irow_i;
 59:   PetscInt        *nrow, *ncol, *w1, *w3, *w4, *rtable, start;
 60:   PetscInt       **sbuf1, m, j, k, l, ct1, **rbuf1, row, proc;
 61:   PetscInt         nrqs, msz, **ptr, *ctr, *pa, *tmp, bsz, nrqr;
 62:   PetscInt         is_no, jmax, **rmap, *rmap_i;
 63:   PetscInt         ctr_j, *sbuf1_j, *rbuf1_i;
 64:   MPI_Request     *s_waits1, *r_waits1, *s_waits2, *r_waits2;
 65:   MPI_Status      *r_status1, *r_status2, *s_status1, *s_status2;
 66:   MPI_Comm         comm;
 67:   PetscScalar    **rbuf2, **sbuf2;
 68:   PetscBool        sorted;

 70:   PetscFunctionBegin;
 71:   PetscCall(PetscObjectGetComm((PetscObject)C, &comm));
 72:   tag0 = ((PetscObject)C)->tag;
 73:   PetscCallMPI(MPI_Comm_rank(comm, &rank));
 74:   PetscCallMPI(MPI_Comm_size(comm, &size));
 75:   m = C->rmap->N;

 77:   /* Get some new tags to keep the communication clean */
 78:   PetscCall(PetscObjectGetNewTag((PetscObject)C, &tag1));

 80:   /* Check if the col indices are sorted */
 81:   for (i = 0; i < ismax; i++) {
 82:     PetscCall(ISSorted(isrow[i], &sorted));
 83:     PetscCheck(sorted, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "ISrow is not sorted");
 84:     PetscCall(ISSorted(iscol[i], &sorted));
 85:     PetscCheck(sorted, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "IScol is not sorted");
 86:   }

 88:   PetscCall(PetscMalloc5(ismax, (PetscInt ***)&irow, ismax, (PetscInt ***)&icol, ismax, &nrow, ismax, &ncol, m, &rtable));
 89:   for (i = 0; i < ismax; i++) {
 90:     PetscCall(ISGetIndices(isrow[i], &irow[i]));
 91:     PetscCall(ISGetIndices(iscol[i], &icol[i]));
 92:     PetscCall(ISGetLocalSize(isrow[i], &nrow[i]));
 93:     PetscCall(ISGetLocalSize(iscol[i], &ncol[i]));
 94:   }

 96:   /* Create hash table for the mapping :row -> proc*/
 97:   for (i = 0, j = 0; i < size; i++) {
 98:     jmax = C->rmap->range[i + 1];
 99:     for (; j < jmax; j++) rtable[j] = i;
100:   }

102:   /* evaluate communication - mesg to who,length of mesg, and buffer space
103:      required. Based on this, buffers are allocated, and data copied into them*/
104:   PetscCall(PetscMalloc3(2 * size, &w1, size, &w3, size, &w4));
105:   PetscCall(PetscArrayzero(w1, size * 2)); /* initialize work vector*/
106:   PetscCall(PetscArrayzero(w3, size));     /* initialize work vector*/
107:   for (i = 0; i < ismax; i++) {
108:     PetscCall(PetscArrayzero(w4, size)); /* initialize work vector*/
109:     jmax   = nrow[i];
110:     irow_i = irow[i];
111:     for (j = 0; j < jmax; j++) {
112:       row  = irow_i[j];
113:       proc = rtable[row];
114:       w4[proc]++;
115:     }
116:     for (j = 0; j < size; j++) {
117:       if (w4[j]) {
118:         w1[2 * j] += w4[j];
119:         w3[j]++;
120:       }
121:     }
122:   }

124:   nrqs         = 0; /* no of outgoing messages */
125:   msz          = 0; /* total mesg length (for all procs) */
126:   w1[2 * rank] = 0; /* no mesg sent to self */
127:   w3[rank]     = 0;
128:   for (i = 0; i < size; i++) {
129:     if (w1[2 * i]) {
130:       w1[2 * i + 1] = 1;
131:       nrqs++;
132:     } /* there exists a message to proc i */
133:   }
134:   PetscCall(PetscMalloc1(nrqs + 1, &pa)); /*(proc -array)*/
135:   for (i = 0, j = 0; i < size; i++) {
136:     if (w1[2 * i]) {
137:       pa[j] = i;
138:       j++;
139:     }
140:   }

142:   /* Each message would have a header = 1 + 2*(no of IS) + data */
143:   for (i = 0; i < nrqs; i++) {
144:     j = pa[i];
145:     w1[2 * j] += w1[2 * j + 1] + 2 * w3[j];
146:     msz += w1[2 * j];
147:   }
148:   /* Do a global reduction to determine how many messages to expect*/
149:   PetscCall(PetscMaxSum(comm, w1, &bsz, &nrqr));

151:   /* Allocate memory for recv buffers . Make sure rbuf1[0] exists by adding 1 to the buffer length */
152:   PetscCall(PetscMalloc1(nrqr + 1, &rbuf1));
153:   PetscCall(PetscMalloc1(nrqr * bsz, &rbuf1[0]));
154:   for (i = 1; i < nrqr; ++i) rbuf1[i] = rbuf1[i - 1] + bsz;

156:   /* Post the receives */
157:   PetscCall(PetscMalloc1(nrqr + 1, &r_waits1));
158:   for (i = 0; i < nrqr; ++i) PetscCallMPI(MPI_Irecv(rbuf1[i], bsz, MPIU_INT, MPI_ANY_SOURCE, tag0, comm, r_waits1 + i));

160:   /* Allocate Memory for outgoing messages */
161:   PetscCall(PetscMalloc4(size, &sbuf1, size, &ptr, 2 * msz, &tmp, size, &ctr));
162:   PetscCall(PetscArrayzero(sbuf1, size));
163:   PetscCall(PetscArrayzero(ptr, size));
164:   {
165:     PetscInt *iptr = tmp, ict = 0;
166:     for (i = 0; i < nrqs; i++) {
167:       j = pa[i];
168:       iptr += ict;
169:       sbuf1[j] = iptr;
170:       ict      = w1[2 * j];
171:     }
172:   }

174:   /* Form the outgoing messages */
175:   /* Initialize the header space */
176:   for (i = 0; i < nrqs; i++) {
177:     j           = pa[i];
178:     sbuf1[j][0] = 0;
179:     PetscCall(PetscArrayzero(sbuf1[j] + 1, 2 * w3[j]));
180:     ptr[j] = sbuf1[j] + 2 * w3[j] + 1;
181:   }

183:   /* Parse the isrow and copy data into outbuf */
184:   for (i = 0; i < ismax; i++) {
185:     PetscCall(PetscArrayzero(ctr, size));
186:     irow_i = irow[i];
187:     jmax   = nrow[i];
188:     for (j = 0; j < jmax; j++) { /* parse the indices of each IS */
189:       row  = irow_i[j];
190:       proc = rtable[row];
191:       if (proc != rank) { /* copy to the outgoing buf*/
192:         ctr[proc]++;
193:         *ptr[proc] = row;
194:         ptr[proc]++;
195:       }
196:     }
197:     /* Update the headers for the current IS */
198:     for (j = 0; j < size; j++) { /* Can Optimise this loop too */
199:       if ((ctr_j = ctr[j])) {
200:         sbuf1_j            = sbuf1[j];
201:         k                  = ++sbuf1_j[0];
202:         sbuf1_j[2 * k]     = ctr_j;
203:         sbuf1_j[2 * k - 1] = i;
204:       }
205:     }
206:   }

208:   /*  Now  post the sends */
209:   PetscCall(PetscMalloc1(nrqs + 1, &s_waits1));
210:   for (i = 0; i < nrqs; ++i) {
211:     j = pa[i];
212:     PetscCallMPI(MPI_Isend(sbuf1[j], w1[2 * j], MPIU_INT, j, tag0, comm, s_waits1 + i));
213:   }

215:   /* Post receives to capture the row_data from other procs */
216:   PetscCall(PetscMalloc1(nrqs + 1, &r_waits2));
217:   PetscCall(PetscMalloc1(nrqs + 1, &rbuf2));
218:   for (i = 0; i < nrqs; i++) {
219:     j     = pa[i];
220:     count = (w1[2 * j] - (2 * sbuf1[j][0] + 1)) * N;
221:     PetscCall(PetscMalloc1(count + 1, &rbuf2[i]));
222:     PetscCallMPI(MPI_Irecv(rbuf2[i], count, MPIU_SCALAR, j, tag1, comm, r_waits2 + i));
223:   }

225:   /* Receive messages(row_nos) and then, pack and send off the rowvalues
226:      to the correct processors */

228:   PetscCall(PetscMalloc1(nrqr + 1, &s_waits2));
229:   PetscCall(PetscMalloc1(nrqr + 1, &r_status1));
230:   PetscCall(PetscMalloc1(nrqr + 1, &sbuf2));

232:   {
233:     PetscScalar *sbuf2_i, *v_start;
234:     PetscInt     s_proc;
235:     for (i = 0; i < nrqr; ++i) {
236:       PetscCallMPI(MPI_Waitany(nrqr, r_waits1, &idex, r_status1 + i));
237:       s_proc  = r_status1[i].MPI_SOURCE; /* send processor */
238:       rbuf1_i = rbuf1[idex];             /* Actual message from s_proc */
239:       /* no of rows = end - start; since start is array idex[], 0idex, whel end
240:          is length of the buffer - which is 1idex */
241:       start = 2 * rbuf1_i[0] + 1;
242:       PetscCallMPI(MPI_Get_count(r_status1 + i, MPIU_INT, &end));
243:       /* allocate memory sufficinet to hold all the row values */
244:       PetscCall(PetscMalloc1((end - start) * N, &sbuf2[idex]));
245:       sbuf2_i = sbuf2[idex];
246:       /* Now pack the data */
247:       for (j = start; j < end; j++) {
248:         row     = rbuf1_i[j] - rstart;
249:         v_start = a->v + row;
250:         for (k = 0; k < N; k++) {
251:           sbuf2_i[0] = v_start[0];
252:           sbuf2_i++;
253:           v_start += a->lda;
254:         }
255:       }
256:       /* Now send off the data */
257:       PetscCallMPI(MPI_Isend(sbuf2[idex], (end - start) * N, MPIU_SCALAR, s_proc, tag1, comm, s_waits2 + i));
258:     }
259:   }
260:   /* End Send-Recv of IS + row_numbers */
261:   PetscCall(PetscFree(r_status1));
262:   PetscCall(PetscFree(r_waits1));
263:   PetscCall(PetscMalloc1(nrqs + 1, &s_status1));
264:   if (nrqs) PetscCallMPI(MPI_Waitall(nrqs, s_waits1, s_status1));
265:   PetscCall(PetscFree(s_status1));
266:   PetscCall(PetscFree(s_waits1));

268:   /* Create the submatrices */
269:   if (scall == MAT_REUSE_MATRIX) {
270:     for (i = 0; i < ismax; i++) {
271:       mat = (Mat_SeqDense *)(submats[i]->data);
272:       PetscCheck(!(submats[i]->rmap->n != nrow[i]) && !(submats[i]->cmap->n != ncol[i]), PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Cannot reuse matrix. wrong size");
273:       PetscCall(PetscArrayzero(mat->v, submats[i]->rmap->n * submats[i]->cmap->n));

275:       submats[i]->factortype = C->factortype;
276:     }
277:   } else {
278:     for (i = 0; i < ismax; i++) {
279:       PetscCall(MatCreate(PETSC_COMM_SELF, submats + i));
280:       PetscCall(MatSetSizes(submats[i], nrow[i], ncol[i], nrow[i], ncol[i]));
281:       PetscCall(MatSetType(submats[i], ((PetscObject)A)->type_name));
282:       PetscCall(MatSeqDenseSetPreallocation(submats[i], NULL));
283:     }
284:   }

286:   /* Assemble the matrices */
287:   {
288:     PetscInt     col;
289:     PetscScalar *imat_v, *mat_v, *imat_vi, *mat_vi;

291:     for (i = 0; i < ismax; i++) {
292:       mat    = (Mat_SeqDense *)submats[i]->data;
293:       mat_v  = a->v;
294:       imat_v = mat->v;
295:       irow_i = irow[i];
296:       m      = nrow[i];
297:       for (j = 0; j < m; j++) {
298:         row  = irow_i[j];
299:         proc = rtable[row];
300:         if (proc == rank) {
301:           row     = row - rstart;
302:           mat_vi  = mat_v + row;
303:           imat_vi = imat_v + j;
304:           for (k = 0; k < ncol[i]; k++) {
305:             col            = icol[i][k];
306:             imat_vi[k * m] = mat_vi[col * a->lda];
307:           }
308:         }
309:       }
310:     }
311:   }

313:   /* Create row map-> This maps c->row to submat->row for each submat*/
314:   /* this is a very expensive operation wrt memory usage */
315:   PetscCall(PetscMalloc1(ismax, &rmap));
316:   PetscCall(PetscCalloc1(ismax * C->rmap->N, &rmap[0]));
317:   for (i = 1; i < ismax; i++) rmap[i] = rmap[i - 1] + C->rmap->N;
318:   for (i = 0; i < ismax; i++) {
319:     rmap_i = rmap[i];
320:     irow_i = irow[i];
321:     jmax   = nrow[i];
322:     for (j = 0; j < jmax; j++) rmap_i[irow_i[j]] = j;
323:   }

325:   /* Now Receive the row_values and assemble the rest of the matrix */
326:   PetscCall(PetscMalloc1(nrqs + 1, &r_status2));
327:   {
328:     PetscInt     is_max, tmp1, col, *sbuf1_i, is_sz;
329:     PetscScalar *rbuf2_i, *imat_v, *imat_vi;

331:     for (tmp1 = 0; tmp1 < nrqs; tmp1++) { /* For each message */
332:       PetscCallMPI(MPI_Waitany(nrqs, r_waits2, &i, r_status2 + tmp1));
333:       /* Now dig out the corresponding sbuf1, which contains the IS data_structure */
334:       sbuf1_i = sbuf1[pa[i]];
335:       is_max  = sbuf1_i[0];
336:       ct1     = 2 * is_max + 1;
337:       rbuf2_i = rbuf2[i];
338:       for (j = 1; j <= is_max; j++) { /* For each IS belonging to the message */
339:         is_no  = sbuf1_i[2 * j - 1];
340:         is_sz  = sbuf1_i[2 * j];
341:         mat    = (Mat_SeqDense *)submats[is_no]->data;
342:         imat_v = mat->v;
343:         rmap_i = rmap[is_no];
344:         m      = nrow[is_no];
345:         for (k = 0; k < is_sz; k++, rbuf2_i += N) { /* For each row */
346:           row = sbuf1_i[ct1];
347:           ct1++;
348:           row     = rmap_i[row];
349:           imat_vi = imat_v + row;
350:           for (l = 0; l < ncol[is_no]; l++) { /* For each col */
351:             col            = icol[is_no][l];
352:             imat_vi[l * m] = rbuf2_i[col];
353:           }
354:         }
355:       }
356:     }
357:   }
358:   /* End Send-Recv of row_values */
359:   PetscCall(PetscFree(r_status2));
360:   PetscCall(PetscFree(r_waits2));
361:   PetscCall(PetscMalloc1(nrqr + 1, &s_status2));
362:   if (nrqr) PetscCallMPI(MPI_Waitall(nrqr, s_waits2, s_status2));
363:   PetscCall(PetscFree(s_status2));
364:   PetscCall(PetscFree(s_waits2));

366:   /* Restore the indices */
367:   for (i = 0; i < ismax; i++) {
368:     PetscCall(ISRestoreIndices(isrow[i], irow + i));
369:     PetscCall(ISRestoreIndices(iscol[i], icol + i));
370:   }

372:   PetscCall(PetscFree5(*(PetscInt ***)&irow, *(PetscInt ***)&icol, nrow, ncol, rtable));
373:   PetscCall(PetscFree3(w1, w3, w4));
374:   PetscCall(PetscFree(pa));

376:   for (i = 0; i < nrqs; ++i) PetscCall(PetscFree(rbuf2[i]));
377:   PetscCall(PetscFree(rbuf2));
378:   PetscCall(PetscFree4(sbuf1, ptr, tmp, ctr));
379:   PetscCall(PetscFree(rbuf1[0]));
380:   PetscCall(PetscFree(rbuf1));

382:   for (i = 0; i < nrqr; ++i) PetscCall(PetscFree(sbuf2[i]));

384:   PetscCall(PetscFree(sbuf2));
385:   PetscCall(PetscFree(rmap[0]));
386:   PetscCall(PetscFree(rmap));

388:   for (i = 0; i < ismax; i++) {
389:     PetscCall(MatAssemblyBegin(submats[i], MAT_FINAL_ASSEMBLY));
390:     PetscCall(MatAssemblyEnd(submats[i], MAT_FINAL_ASSEMBLY));
391:   }
392:   PetscFunctionReturn(PETSC_SUCCESS);
393: }

395: PETSC_INTERN PetscErrorCode MatScale_MPIDense(Mat inA, PetscScalar alpha)
396: {
397:   Mat_MPIDense *A = (Mat_MPIDense *)inA->data;

399:   PetscFunctionBegin;
400:   PetscCall(MatScale(A->A, alpha));
401:   PetscFunctionReturn(PETSC_SUCCESS);
402: }