1 /* 2 Routines to compute overlapping regions of a parallel MPI matrix 3 and to find submatrices that were shared across processors. 4 */ 5 #include <../src/mat/impls/aij/seq/aij.h> 6 #include <../src/mat/impls/aij/mpi/mpiaij.h> 7 #include <petscbt.h> 8 #include <petscsf.h> 9 10 static PetscErrorCode MatIncreaseOverlap_MPIAIJ_Once(Mat,PetscInt,IS*); 11 static PetscErrorCode MatIncreaseOverlap_MPIAIJ_Local(Mat,PetscInt,char**,PetscInt*,PetscInt**,PetscTable*); 12 static PetscErrorCode MatIncreaseOverlap_MPIAIJ_Receive(Mat,PetscInt,PetscInt**,PetscInt**,PetscInt*); 13 extern PetscErrorCode MatGetRow_MPIAIJ(Mat,PetscInt,PetscInt*,PetscInt**,PetscScalar**); 14 extern PetscErrorCode MatRestoreRow_MPIAIJ(Mat,PetscInt,PetscInt*,PetscInt**,PetscScalar**); 15 16 static PetscErrorCode MatIncreaseOverlap_MPIAIJ_Once_Scalable(Mat,PetscInt,IS*); 17 static PetscErrorCode MatIncreaseOverlap_MPIAIJ_Local_Scalable(Mat,PetscInt,IS*); 18 static PetscErrorCode MatIncreaseOverlap_MPIAIJ_Send_Scalable(Mat,PetscInt,PetscMPIInt,PetscMPIInt *,PetscInt *, PetscInt *,PetscInt **,PetscInt **); 19 static PetscErrorCode MatIncreaseOverlap_MPIAIJ_Receive_Scalable(Mat,PetscInt,IS*,PetscInt,PetscInt *); 20 21 22 PetscErrorCode MatIncreaseOverlap_MPIAIJ(Mat C,PetscInt imax,IS is[],PetscInt ov) 23 { 24 PetscErrorCode ierr; 25 PetscInt i; 26 27 PetscFunctionBegin; 28 if (ov < 0) SETERRQ(PetscObjectComm((PetscObject)C),PETSC_ERR_ARG_OUTOFRANGE,"Negative overlap specified"); 29 for (i=0; i<ov; ++i) { 30 ierr = MatIncreaseOverlap_MPIAIJ_Once(C,imax,is);CHKERRQ(ierr); 31 } 32 PetscFunctionReturn(0); 33 } 34 35 PetscErrorCode MatIncreaseOverlap_MPIAIJ_Scalable(Mat C,PetscInt imax,IS is[],PetscInt ov) 36 { 37 PetscErrorCode ierr; 38 PetscInt i; 39 40 PetscFunctionBegin; 41 if (ov < 0) SETERRQ(PetscObjectComm((PetscObject)C),PETSC_ERR_ARG_OUTOFRANGE,"Negative overlap specified"); 42 for (i=0; i<ov; ++i) { 43 ierr = MatIncreaseOverlap_MPIAIJ_Once_Scalable(C,imax,is);CHKERRQ(ierr); 44 } 45 PetscFunctionReturn(0); 46 } 47 48 49 static PetscErrorCode MatIncreaseOverlap_MPIAIJ_Once_Scalable(Mat mat,PetscInt nidx,IS is[]) 50 { 51 PetscErrorCode ierr; 52 MPI_Comm comm; 53 PetscInt *length,length_i,tlength,*remoterows,nrrows,reducednrrows,*rrow_ranks,*rrow_isids,i,j,owner; 54 PetscInt *tosizes,*tosizes_temp,*toffsets,*fromsizes,*todata,*fromdata; 55 PetscInt nrecvrows,*sbsizes = 0,*sbdata = 0; 56 const PetscInt *indices_i,**indices; 57 PetscLayout rmap; 58 PetscMPIInt rank,size,*toranks,*fromranks,nto,nfrom; 59 PetscSF sf; 60 PetscSFNode *remote; 61 62 PetscFunctionBegin; 63 ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 64 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 65 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 66 /* get row map to determine where rows should be going */ 67 ierr = MatGetLayouts(mat,&rmap,NULL);CHKERRQ(ierr); 68 /* retrieve IS data and put all together so that we 69 * can optimize communication 70 * */ 71 ierr = PetscCalloc2(nidx,(PetscInt ***)&indices,nidx,&length);CHKERRQ(ierr); 72 for (i=0,tlength=0; i<nidx; i++){ 73 ierr = ISGetLocalSize(is[i],&length[i]);CHKERRQ(ierr); 74 tlength += length[i]; 75 ierr = ISGetIndices(is[i],&indices[i]);CHKERRQ(ierr); 76 } 77 /* find these rows on remote processors */ 78 ierr = PetscCalloc3(tlength,&remoterows,tlength,&rrow_ranks,tlength,&rrow_isids);CHKERRQ(ierr); 79 ierr = PetscCalloc3(size,&toranks,2*size,&tosizes,size,&tosizes_temp);CHKERRQ(ierr); 80 nrrows = 0; 81 for (i=0; i<nidx; i++){ 82 length_i = length[i]; 83 indices_i = indices[i]; 84 for (j=0; j<length_i; j++){ 85 owner = -1; 86 ierr = PetscLayoutFindOwner(rmap,indices_i[j],&owner);CHKERRQ(ierr); 87 /* remote processors */ 88 if (owner != rank){ 89 tosizes_temp[owner]++; /* number of rows to owner */ 90 rrow_ranks[nrrows] = owner; /* processor */ 91 rrow_isids[nrrows] = i; /* is id */ 92 remoterows[nrrows++] = indices_i[j]; /* row */ 93 } 94 } 95 ierr = ISRestoreIndices(is[i],&indices[i]);CHKERRQ(ierr); 96 } 97 ierr = PetscFree2(indices,length);CHKERRQ(ierr); 98 /* test if we need to exchange messages 99 * generally speaking, we do not need to exchange 100 * data when overlap is 1 101 * */ 102 ierr = MPIU_Allreduce(&nrrows,&reducednrrows,1,MPIU_INT,MPIU_MAX,comm);CHKERRQ(ierr); 103 /* we do not have any messages 104 * It usually corresponds to overlap 1 105 * */ 106 if (!reducednrrows){ 107 ierr = PetscFree3(toranks,tosizes,tosizes_temp);CHKERRQ(ierr); 108 ierr = PetscFree3(remoterows,rrow_ranks,rrow_isids);CHKERRQ(ierr); 109 ierr = MatIncreaseOverlap_MPIAIJ_Local_Scalable(mat,nidx,is);CHKERRQ(ierr); 110 PetscFunctionReturn(0); 111 } 112 nto = 0; 113 /* send sizes and ranks for building a two-sided communcation */ 114 for (i=0; i<size; i++){ 115 if (tosizes_temp[i]){ 116 tosizes[nto*2] = tosizes_temp[i]*2; /* size */ 117 tosizes_temp[i] = nto; /* a map from processor to index */ 118 toranks[nto++] = i; /* processor */ 119 } 120 } 121 ierr = PetscCalloc1(nto+1,&toffsets);CHKERRQ(ierr); 122 for (i=0; i<nto; i++){ 123 toffsets[i+1] = toffsets[i]+tosizes[2*i]; /* offsets */ 124 tosizes[2*i+1] = toffsets[i]; /* offsets to send */ 125 } 126 /* send information to other processors */ 127 ierr = PetscCommBuildTwoSided(comm,2,MPIU_INT,nto,toranks,tosizes,&nfrom,&fromranks,&fromsizes);CHKERRQ(ierr); 128 nrecvrows = 0; 129 for (i=0; i<nfrom; i++) nrecvrows += fromsizes[2*i]; 130 ierr = PetscMalloc1(nrecvrows,&remote);CHKERRQ(ierr); 131 nrecvrows = 0; 132 for (i=0; i<nfrom; i++){ 133 for (j=0; j<fromsizes[2*i]; j++){ 134 remote[nrecvrows].rank = fromranks[i]; 135 remote[nrecvrows++].index = fromsizes[2*i+1]+j; 136 } 137 } 138 ierr = PetscSFCreate(comm,&sf);CHKERRQ(ierr); 139 ierr = PetscSFSetGraph(sf,nrecvrows,nrecvrows,NULL,PETSC_OWN_POINTER,remote,PETSC_OWN_POINTER);CHKERRQ(ierr); 140 /* use two-sided communication by default since OPENMPI has some bugs for one-sided one */ 141 ierr = PetscSFSetType(sf,PETSCSFBASIC);CHKERRQ(ierr); 142 ierr = PetscSFSetFromOptions(sf);CHKERRQ(ierr); 143 /* message pair <no of is, row> */ 144 ierr = PetscCalloc2(2*nrrows,&todata,nrecvrows,&fromdata);CHKERRQ(ierr); 145 for (i=0; i<nrrows; i++){ 146 owner = rrow_ranks[i]; /* processor */ 147 j = tosizes_temp[owner]; /* index */ 148 todata[toffsets[j]++] = rrow_isids[i]; 149 todata[toffsets[j]++] = remoterows[i]; 150 } 151 ierr = PetscFree3(toranks,tosizes,tosizes_temp);CHKERRQ(ierr); 152 ierr = PetscFree3(remoterows,rrow_ranks,rrow_isids);CHKERRQ(ierr); 153 ierr = PetscFree(toffsets);CHKERRQ(ierr); 154 ierr = PetscSFBcastBegin(sf,MPIU_INT,todata,fromdata);CHKERRQ(ierr); 155 ierr = PetscSFBcastEnd(sf,MPIU_INT,todata,fromdata);CHKERRQ(ierr); 156 ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 157 /* send rows belonging to the remote so that then we could get the overlapping data back */ 158 ierr = MatIncreaseOverlap_MPIAIJ_Send_Scalable(mat,nidx,nfrom,fromranks,fromsizes,fromdata,&sbsizes,&sbdata);CHKERRQ(ierr); 159 ierr = PetscFree2(todata,fromdata);CHKERRQ(ierr); 160 ierr = PetscFree(fromsizes);CHKERRQ(ierr); 161 ierr = PetscCommBuildTwoSided(comm,2,MPIU_INT,nfrom,fromranks,sbsizes,&nto,&toranks,&tosizes);CHKERRQ(ierr); 162 ierr = PetscFree(fromranks);CHKERRQ(ierr); 163 nrecvrows = 0; 164 for (i=0; i<nto; i++) nrecvrows += tosizes[2*i]; 165 ierr = PetscCalloc1(nrecvrows,&todata);CHKERRQ(ierr); 166 ierr = PetscMalloc1(nrecvrows,&remote);CHKERRQ(ierr); 167 nrecvrows = 0; 168 for (i=0; i<nto; i++){ 169 for (j=0; j<tosizes[2*i]; j++){ 170 remote[nrecvrows].rank = toranks[i]; 171 remote[nrecvrows++].index = tosizes[2*i+1]+j; 172 } 173 } 174 ierr = PetscSFCreate(comm,&sf);CHKERRQ(ierr); 175 ierr = PetscSFSetGraph(sf,nrecvrows,nrecvrows,NULL,PETSC_OWN_POINTER,remote,PETSC_OWN_POINTER);CHKERRQ(ierr); 176 /* use two-sided communication by default since OPENMPI has some bugs for one-sided one */ 177 ierr = PetscSFSetType(sf,PETSCSFBASIC);CHKERRQ(ierr); 178 ierr = PetscSFSetFromOptions(sf);CHKERRQ(ierr); 179 /* overlap communication and computation */ 180 ierr = PetscSFBcastBegin(sf,MPIU_INT,sbdata,todata);CHKERRQ(ierr); 181 ierr = MatIncreaseOverlap_MPIAIJ_Local_Scalable(mat,nidx,is);CHKERRQ(ierr); 182 ierr = PetscSFBcastEnd(sf,MPIU_INT,sbdata,todata);CHKERRQ(ierr); 183 ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 184 ierr = PetscFree2(sbdata,sbsizes);CHKERRQ(ierr); 185 ierr = MatIncreaseOverlap_MPIAIJ_Receive_Scalable(mat,nidx,is,nrecvrows,todata);CHKERRQ(ierr); 186 ierr = PetscFree(toranks);CHKERRQ(ierr); 187 ierr = PetscFree(tosizes);CHKERRQ(ierr); 188 ierr = PetscFree(todata);CHKERRQ(ierr); 189 PetscFunctionReturn(0); 190 } 191 192 static PetscErrorCode MatIncreaseOverlap_MPIAIJ_Receive_Scalable(Mat mat,PetscInt nidx, IS is[], PetscInt nrecvs, PetscInt *recvdata) 193 { 194 PetscInt *isz,isz_i,i,j,is_id, data_size; 195 PetscInt col,lsize,max_lsize,*indices_temp, *indices_i; 196 const PetscInt *indices_i_temp; 197 PetscErrorCode ierr; 198 199 PetscFunctionBegin; 200 max_lsize = 0; 201 ierr = PetscMalloc1(nidx,&isz);CHKERRQ(ierr); 202 for (i=0; i<nidx; i++){ 203 ierr = ISGetLocalSize(is[i],&lsize);CHKERRQ(ierr); 204 max_lsize = lsize>max_lsize ? lsize:max_lsize; 205 isz[i] = lsize; 206 } 207 ierr = PetscMalloc1((max_lsize+nrecvs)*nidx,&indices_temp);CHKERRQ(ierr); 208 for (i=0; i<nidx; i++){ 209 ierr = ISGetIndices(is[i],&indices_i_temp);CHKERRQ(ierr); 210 ierr = PetscMemcpy(indices_temp+i*(max_lsize+nrecvs),indices_i_temp, sizeof(PetscInt)*isz[i]);CHKERRQ(ierr); 211 ierr = ISRestoreIndices(is[i],&indices_i_temp);CHKERRQ(ierr); 212 ierr = ISDestroy(&is[i]);CHKERRQ(ierr); 213 } 214 /* retrieve information to get row id and its overlap */ 215 for (i=0; i<nrecvs; ){ 216 is_id = recvdata[i++]; 217 data_size = recvdata[i++]; 218 indices_i = indices_temp+(max_lsize+nrecvs)*is_id; 219 isz_i = isz[is_id]; 220 for (j=0; j< data_size; j++){ 221 col = recvdata[i++]; 222 indices_i[isz_i++] = col; 223 } 224 isz[is_id] = isz_i; 225 } 226 /* remove duplicate entities */ 227 for (i=0; i<nidx; i++){ 228 indices_i = indices_temp+(max_lsize+nrecvs)*i; 229 isz_i = isz[i]; 230 ierr = PetscSortRemoveDupsInt(&isz_i,indices_i);CHKERRQ(ierr); 231 ierr = ISCreateGeneral(PETSC_COMM_SELF,isz_i,indices_i,PETSC_COPY_VALUES,&is[i]);CHKERRQ(ierr); 232 } 233 ierr = PetscFree(isz);CHKERRQ(ierr); 234 ierr = PetscFree(indices_temp);CHKERRQ(ierr); 235 PetscFunctionReturn(0); 236 } 237 238 static PetscErrorCode MatIncreaseOverlap_MPIAIJ_Send_Scalable(Mat mat,PetscInt nidx, PetscMPIInt nfrom,PetscMPIInt *fromranks,PetscInt *fromsizes, PetscInt *fromrows, PetscInt **sbrowsizes, PetscInt **sbrows) 239 { 240 PetscLayout rmap,cmap; 241 PetscInt i,j,k,l,*rows_i,*rows_data_ptr,**rows_data,max_fszs,rows_pos,*rows_pos_i; 242 PetscInt is_id,tnz,an,bn,rstart,cstart,row,start,end,col,totalrows,*sbdata; 243 PetscInt *indv_counts,indvc_ij,*sbsizes,*indices_tmp,*offsets; 244 const PetscInt *gcols,*ai,*aj,*bi,*bj; 245 Mat amat,bmat; 246 PetscMPIInt rank; 247 PetscBool done; 248 MPI_Comm comm; 249 PetscErrorCode ierr; 250 251 PetscFunctionBegin; 252 ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 253 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 254 ierr = MatMPIAIJGetSeqAIJ(mat,&amat,&bmat,&gcols);CHKERRQ(ierr); 255 /* Even if the mat is symmetric, we still assume it is not symmetric */ 256 ierr = MatGetRowIJ(amat,0,PETSC_FALSE,PETSC_FALSE,&an,&ai,&aj,&done);CHKERRQ(ierr); 257 if (!done) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"can not get row IJ \n"); 258 ierr = MatGetRowIJ(bmat,0,PETSC_FALSE,PETSC_FALSE,&bn,&bi,&bj,&done);CHKERRQ(ierr); 259 if (!done) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"can not get row IJ \n"); 260 /* total number of nonzero values is used to estimate the memory usage in the next step */ 261 tnz = ai[an]+bi[bn]; 262 ierr = MatGetLayouts(mat,&rmap,&cmap);CHKERRQ(ierr); 263 ierr = PetscLayoutGetRange(rmap,&rstart,NULL);CHKERRQ(ierr); 264 ierr = PetscLayoutGetRange(cmap,&cstart,NULL);CHKERRQ(ierr); 265 /* to find the longest message */ 266 max_fszs = 0; 267 for (i=0; i<nfrom; i++) max_fszs = fromsizes[2*i]>max_fszs ? fromsizes[2*i]:max_fszs; 268 /* better way to estimate number of nonzero in the mat??? */ 269 ierr = PetscCalloc5(max_fszs*nidx,&rows_data_ptr,nidx,&rows_data,nidx,&rows_pos_i,nfrom*nidx,&indv_counts,tnz,&indices_tmp);CHKERRQ(ierr); 270 for (i=0; i<nidx; i++) rows_data[i] = rows_data_ptr+max_fszs*i; 271 rows_pos = 0; 272 totalrows = 0; 273 for (i=0; i<nfrom; i++){ 274 ierr = PetscMemzero(rows_pos_i,sizeof(PetscInt)*nidx);CHKERRQ(ierr); 275 /* group data together */ 276 for (j=0; j<fromsizes[2*i]; j+=2){ 277 is_id = fromrows[rows_pos++];/* no of is */ 278 rows_i = rows_data[is_id]; 279 rows_i[rows_pos_i[is_id]++] = fromrows[rows_pos++];/* row */ 280 } 281 /* estimate a space to avoid multiple allocations */ 282 for (j=0; j<nidx; j++){ 283 indvc_ij = 0; 284 rows_i = rows_data[j]; 285 for (l=0; l<rows_pos_i[j]; l++){ 286 row = rows_i[l]-rstart; 287 start = ai[row]; 288 end = ai[row+1]; 289 for (k=start; k<end; k++){ /* Amat */ 290 col = aj[k] + cstart; 291 indices_tmp[indvc_ij++] = col;/* do not count the rows from the original rank */ 292 } 293 start = bi[row]; 294 end = bi[row+1]; 295 for (k=start; k<end; k++) { /* Bmat */ 296 col = gcols[bj[k]]; 297 indices_tmp[indvc_ij++] = col; 298 } 299 } 300 ierr = PetscSortRemoveDupsInt(&indvc_ij,indices_tmp);CHKERRQ(ierr); 301 indv_counts[i*nidx+j] = indvc_ij; 302 totalrows += indvc_ij; 303 } 304 } 305 /* message triple <no of is, number of rows, rows> */ 306 ierr = PetscCalloc2(totalrows+nidx*nfrom*2,&sbdata,2*nfrom,&sbsizes);CHKERRQ(ierr); 307 totalrows = 0; 308 rows_pos = 0; 309 /* use this code again */ 310 for (i=0;i<nfrom;i++){ 311 ierr = PetscMemzero(rows_pos_i,sizeof(PetscInt)*nidx);CHKERRQ(ierr); 312 for (j=0; j<fromsizes[2*i]; j+=2){ 313 is_id = fromrows[rows_pos++]; 314 rows_i = rows_data[is_id]; 315 rows_i[rows_pos_i[is_id]++] = fromrows[rows_pos++]; 316 } 317 /* add data */ 318 for (j=0; j<nidx; j++){ 319 if (!indv_counts[i*nidx+j]) continue; 320 indvc_ij = 0; 321 sbdata[totalrows++] = j; 322 sbdata[totalrows++] = indv_counts[i*nidx+j]; 323 sbsizes[2*i] += 2; 324 rows_i = rows_data[j]; 325 for (l=0; l<rows_pos_i[j]; l++){ 326 row = rows_i[l]-rstart; 327 start = ai[row]; 328 end = ai[row+1]; 329 for (k=start; k<end; k++){ /* Amat */ 330 col = aj[k] + cstart; 331 indices_tmp[indvc_ij++] = col; 332 } 333 start = bi[row]; 334 end = bi[row+1]; 335 for (k=start; k<end; k++) { /* Bmat */ 336 col = gcols[bj[k]]; 337 indices_tmp[indvc_ij++] = col; 338 } 339 } 340 ierr = PetscSortRemoveDupsInt(&indvc_ij,indices_tmp);CHKERRQ(ierr); 341 sbsizes[2*i] += indvc_ij; 342 ierr = PetscMemcpy(sbdata+totalrows,indices_tmp,sizeof(PetscInt)*indvc_ij);CHKERRQ(ierr); 343 totalrows += indvc_ij; 344 } 345 } 346 ierr = PetscCalloc1(nfrom+1,&offsets);CHKERRQ(ierr); 347 for (i=0; i<nfrom; i++){ 348 offsets[i+1] = offsets[i] + sbsizes[2*i]; 349 sbsizes[2*i+1] = offsets[i]; 350 } 351 ierr = PetscFree(offsets);CHKERRQ(ierr); 352 if (sbrowsizes) *sbrowsizes = sbsizes; 353 if (sbrows) *sbrows = sbdata; 354 ierr = PetscFree5(rows_data_ptr,rows_data,rows_pos_i,indv_counts,indices_tmp);CHKERRQ(ierr); 355 ierr = MatRestoreRowIJ(amat,0,PETSC_FALSE,PETSC_FALSE,&an,&ai,&aj,&done);CHKERRQ(ierr); 356 ierr = MatRestoreRowIJ(bmat,0,PETSC_FALSE,PETSC_FALSE,&bn,&bi,&bj,&done);CHKERRQ(ierr); 357 PetscFunctionReturn(0); 358 } 359 360 static PetscErrorCode MatIncreaseOverlap_MPIAIJ_Local_Scalable(Mat mat,PetscInt nidx, IS is[]) 361 { 362 const PetscInt *gcols,*ai,*aj,*bi,*bj, *indices; 363 PetscInt tnz,an,bn,i,j,row,start,end,rstart,cstart,col,k,*indices_temp; 364 PetscInt lsize,lsize_tmp,owner; 365 PetscMPIInt rank; 366 Mat amat,bmat; 367 PetscBool done; 368 PetscLayout cmap,rmap; 369 MPI_Comm comm; 370 PetscErrorCode ierr; 371 372 PetscFunctionBegin; 373 ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 374 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 375 ierr = MatMPIAIJGetSeqAIJ(mat,&amat,&bmat,&gcols);CHKERRQ(ierr); 376 ierr = MatGetRowIJ(amat,0,PETSC_FALSE,PETSC_FALSE,&an,&ai,&aj,&done);CHKERRQ(ierr); 377 if (!done) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"can not get row IJ \n"); 378 ierr = MatGetRowIJ(bmat,0,PETSC_FALSE,PETSC_FALSE,&bn,&bi,&bj,&done);CHKERRQ(ierr); 379 if (!done) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"can not get row IJ \n"); 380 /* is it a safe way to compute number of nonzero values ? */ 381 tnz = ai[an]+bi[bn]; 382 ierr = MatGetLayouts(mat,&rmap,&cmap);CHKERRQ(ierr); 383 ierr = PetscLayoutGetRange(rmap,&rstart,NULL);CHKERRQ(ierr); 384 ierr = PetscLayoutGetRange(cmap,&cstart,NULL);CHKERRQ(ierr); 385 /* it is a better way to estimate memory than the old implementation 386 * where global size of matrix is used 387 * */ 388 ierr = PetscMalloc1(tnz,&indices_temp);CHKERRQ(ierr); 389 for (i=0; i<nidx; i++) { 390 ierr = ISGetLocalSize(is[i],&lsize);CHKERRQ(ierr); 391 ierr = ISGetIndices(is[i],&indices);CHKERRQ(ierr); 392 lsize_tmp = 0; 393 for (j=0; j<lsize; j++) { 394 owner = -1; 395 row = indices[j]; 396 ierr = PetscLayoutFindOwner(rmap,row,&owner);CHKERRQ(ierr); 397 if (owner != rank) continue; 398 /* local number */ 399 row -= rstart; 400 start = ai[row]; 401 end = ai[row+1]; 402 for (k=start; k<end; k++) { /* Amat */ 403 col = aj[k] + cstart; 404 indices_temp[lsize_tmp++] = col; 405 } 406 start = bi[row]; 407 end = bi[row+1]; 408 for (k=start; k<end; k++) { /* Bmat */ 409 col = gcols[bj[k]]; 410 indices_temp[lsize_tmp++] = col; 411 } 412 } 413 ierr = ISRestoreIndices(is[i],&indices);CHKERRQ(ierr); 414 ierr = ISDestroy(&is[i]);CHKERRQ(ierr); 415 ierr = PetscSortRemoveDupsInt(&lsize_tmp,indices_temp);CHKERRQ(ierr); 416 ierr = ISCreateGeneral(PETSC_COMM_SELF,lsize_tmp,indices_temp,PETSC_COPY_VALUES,&is[i]);CHKERRQ(ierr); 417 } 418 ierr = PetscFree(indices_temp);CHKERRQ(ierr); 419 ierr = MatRestoreRowIJ(amat,0,PETSC_FALSE,PETSC_FALSE,&an,&ai,&aj,&done);CHKERRQ(ierr); 420 ierr = MatRestoreRowIJ(bmat,0,PETSC_FALSE,PETSC_FALSE,&bn,&bi,&bj,&done);CHKERRQ(ierr); 421 PetscFunctionReturn(0); 422 } 423 424 425 /* 426 Sample message format: 427 If a processor A wants processor B to process some elements corresponding 428 to index sets is[1],is[5] 429 mesg [0] = 2 (no of index sets in the mesg) 430 ----------- 431 mesg [1] = 1 => is[1] 432 mesg [2] = sizeof(is[1]); 433 ----------- 434 mesg [3] = 5 => is[5] 435 mesg [4] = sizeof(is[5]); 436 ----------- 437 mesg [5] 438 mesg [n] datas[1] 439 ----------- 440 mesg[n+1] 441 mesg[m] data(is[5]) 442 ----------- 443 444 Notes: 445 nrqs - no of requests sent (or to be sent out) 446 nrqr - no of requests recieved (which have to be or which have been processed 447 */ 448 static PetscErrorCode MatIncreaseOverlap_MPIAIJ_Once(Mat C,PetscInt imax,IS is[]) 449 { 450 Mat_MPIAIJ *c = (Mat_MPIAIJ*)C->data; 451 PetscMPIInt *w1,*w2,nrqr,*w3,*w4,*onodes1,*olengths1,*onodes2,*olengths2; 452 const PetscInt **idx,*idx_i; 453 PetscInt *n,**data,len; 454 #if defined(PETSC_USE_CTABLE) 455 PetscTable *table_data,table_data_i; 456 PetscInt *tdata,tcount,tcount_max; 457 #else 458 PetscInt *data_i,*d_p; 459 #endif 460 PetscErrorCode ierr; 461 PetscMPIInt size,rank,tag1,tag2; 462 PetscInt M,i,j,k,**rbuf,row,proc = 0,nrqs,msz,**outdat,**ptr; 463 PetscInt *ctr,*pa,*tmp,*isz,*isz1,**xdata,**rbuf2; 464 PetscBT *table; 465 MPI_Comm comm; 466 MPI_Request *s_waits1,*r_waits1,*s_waits2,*r_waits2; 467 MPI_Status *s_status,*recv_status; 468 char *t_p; 469 470 PetscFunctionBegin; 471 ierr = PetscObjectGetComm((PetscObject)C,&comm);CHKERRQ(ierr); 472 size = c->size; 473 rank = c->rank; 474 M = C->rmap->N; 475 476 ierr = PetscObjectGetNewTag((PetscObject)C,&tag1);CHKERRQ(ierr); 477 ierr = PetscObjectGetNewTag((PetscObject)C,&tag2);CHKERRQ(ierr); 478 479 ierr = PetscMalloc2(imax,&idx,imax,&n);CHKERRQ(ierr); 480 481 for (i=0; i<imax; i++) { 482 ierr = ISGetIndices(is[i],&idx[i]);CHKERRQ(ierr); 483 ierr = ISGetLocalSize(is[i],&n[i]);CHKERRQ(ierr); 484 } 485 486 /* evaluate communication - mesg to who,length of mesg, and buffer space 487 required. Based on this, buffers are allocated, and data copied into them */ 488 ierr = PetscMalloc4(size,&w1,size,&w2,size,&w3,size,&w4);CHKERRQ(ierr); 489 ierr = PetscMemzero(w1,size*sizeof(PetscMPIInt));CHKERRQ(ierr); /* initialise work vector*/ 490 ierr = PetscMemzero(w2,size*sizeof(PetscMPIInt));CHKERRQ(ierr); /* initialise work vector*/ 491 ierr = PetscMemzero(w3,size*sizeof(PetscMPIInt));CHKERRQ(ierr); /* initialise work vector*/ 492 for (i=0; i<imax; i++) { 493 ierr = PetscMemzero(w4,size*sizeof(PetscMPIInt));CHKERRQ(ierr); /* initialise work vector*/ 494 idx_i = idx[i]; 495 len = n[i]; 496 for (j=0; j<len; j++) { 497 row = idx_i[j]; 498 if (row < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Index set cannot have negative entries"); 499 ierr = PetscLayoutFindOwner(C->rmap,row,&proc);CHKERRQ(ierr); 500 w4[proc]++; 501 } 502 for (j=0; j<size; j++) { 503 if (w4[j]) { w1[j] += w4[j]; w3[j]++;} 504 } 505 } 506 507 nrqs = 0; /* no of outgoing messages */ 508 msz = 0; /* total mesg length (for all proc */ 509 w1[rank] = 0; /* no mesg sent to intself */ 510 w3[rank] = 0; 511 for (i=0; i<size; i++) { 512 if (w1[i]) {w2[i] = 1; nrqs++;} /* there exists a message to proc i */ 513 } 514 /* pa - is list of processors to communicate with */ 515 ierr = PetscMalloc1(nrqs+1,&pa);CHKERRQ(ierr); 516 for (i=0,j=0; i<size; i++) { 517 if (w1[i]) {pa[j] = i; j++;} 518 } 519 520 /* Each message would have a header = 1 + 2*(no of IS) + data */ 521 for (i=0; i<nrqs; i++) { 522 j = pa[i]; 523 w1[j] += w2[j] + 2*w3[j]; 524 msz += w1[j]; 525 } 526 527 /* Determine the number of messages to expect, their lengths, from from-ids */ 528 ierr = PetscGatherNumberOfMessages(comm,w2,w1,&nrqr);CHKERRQ(ierr); 529 ierr = PetscGatherMessageLengths(comm,nrqs,nrqr,w1,&onodes1,&olengths1);CHKERRQ(ierr); 530 531 /* Now post the Irecvs corresponding to these messages */ 532 ierr = PetscPostIrecvInt(comm,tag1,nrqr,onodes1,olengths1,&rbuf,&r_waits1);CHKERRQ(ierr); 533 534 /* Allocate Memory for outgoing messages */ 535 ierr = PetscMalloc4(size,&outdat,size,&ptr,msz,&tmp,size,&ctr);CHKERRQ(ierr); 536 ierr = PetscMemzero(outdat,size*sizeof(PetscInt*));CHKERRQ(ierr); 537 ierr = PetscMemzero(ptr,size*sizeof(PetscInt*));CHKERRQ(ierr); 538 539 { 540 PetscInt *iptr = tmp,ict = 0; 541 for (i=0; i<nrqs; i++) { 542 j = pa[i]; 543 iptr += ict; 544 outdat[j] = iptr; 545 ict = w1[j]; 546 } 547 } 548 549 /* Form the outgoing messages */ 550 /* plug in the headers */ 551 for (i=0; i<nrqs; i++) { 552 j = pa[i]; 553 outdat[j][0] = 0; 554 ierr = PetscMemzero(outdat[j]+1,2*w3[j]*sizeof(PetscInt));CHKERRQ(ierr); 555 ptr[j] = outdat[j] + 2*w3[j] + 1; 556 } 557 558 /* Memory for doing local proc's work */ 559 { 560 PetscInt M_BPB_imax = 0; 561 #if defined(PETSC_USE_CTABLE) 562 ierr = PetscIntMultError((M/PETSC_BITS_PER_BYTE+1),imax, &M_BPB_imax);CHKERRQ(ierr); 563 ierr = PetscMalloc1(imax,&table_data);CHKERRQ(ierr); 564 for (i=0; i<imax; i++) { 565 ierr = PetscTableCreate(n[i]+1,M+1,&table_data[i]);CHKERRQ(ierr); 566 } 567 ierr = PetscCalloc4(imax,&table, imax,&data, imax,&isz, M_BPB_imax,&t_p);CHKERRQ(ierr); 568 for (i=0; i<imax; i++) { 569 table[i] = t_p + (M/PETSC_BITS_PER_BYTE+1)*i; 570 } 571 #else 572 PetscInt Mimax = 0; 573 ierr = PetscIntMultError(M,imax, &Mimax);CHKERRQ(ierr); 574 ierr = PetscIntMultError((M/PETSC_BITS_PER_BYTE+1),imax, &M_BPB_imax);CHKERRQ(ierr); 575 ierr = PetscCalloc5(imax,&table, imax,&data, imax,&isz, Mimax,&d_p, M_BPB_imax,&t_p);CHKERRQ(ierr); 576 for (i=0; i<imax; i++) { 577 table[i] = t_p + (M/PETSC_BITS_PER_BYTE+1)*i; 578 data[i] = d_p + M*i; 579 } 580 #endif 581 } 582 583 /* Parse the IS and update local tables and the outgoing buf with the data */ 584 { 585 PetscInt n_i,isz_i,*outdat_j,ctr_j; 586 PetscBT table_i; 587 588 for (i=0; i<imax; i++) { 589 ierr = PetscMemzero(ctr,size*sizeof(PetscInt));CHKERRQ(ierr); 590 n_i = n[i]; 591 table_i = table[i]; 592 idx_i = idx[i]; 593 #if defined(PETSC_USE_CTABLE) 594 table_data_i = table_data[i]; 595 #else 596 data_i = data[i]; 597 #endif 598 isz_i = isz[i]; 599 for (j=0; j<n_i; j++) { /* parse the indices of each IS */ 600 row = idx_i[j]; 601 ierr = PetscLayoutFindOwner(C->rmap,row,&proc);CHKERRQ(ierr); 602 if (proc != rank) { /* copy to the outgoing buffer */ 603 ctr[proc]++; 604 *ptr[proc] = row; 605 ptr[proc]++; 606 } else if (!PetscBTLookupSet(table_i,row)) { 607 #if defined(PETSC_USE_CTABLE) 608 ierr = PetscTableAdd(table_data_i,row+1,isz_i+1,INSERT_VALUES);CHKERRQ(ierr); 609 #else 610 data_i[isz_i] = row; /* Update the local table */ 611 #endif 612 isz_i++; 613 } 614 } 615 /* Update the headers for the current IS */ 616 for (j=0; j<size; j++) { /* Can Optimise this loop by using pa[] */ 617 if ((ctr_j = ctr[j])) { 618 outdat_j = outdat[j]; 619 k = ++outdat_j[0]; 620 outdat_j[2*k] = ctr_j; 621 outdat_j[2*k-1] = i; 622 } 623 } 624 isz[i] = isz_i; 625 } 626 } 627 628 /* Now post the sends */ 629 ierr = PetscMalloc1(nrqs+1,&s_waits1);CHKERRQ(ierr); 630 for (i=0; i<nrqs; ++i) { 631 j = pa[i]; 632 ierr = MPI_Isend(outdat[j],w1[j],MPIU_INT,j,tag1,comm,s_waits1+i);CHKERRQ(ierr); 633 } 634 635 /* No longer need the original indices */ 636 for (i=0; i<imax; ++i) { 637 ierr = ISRestoreIndices(is[i],idx+i);CHKERRQ(ierr); 638 } 639 ierr = PetscFree2(idx,n);CHKERRQ(ierr); 640 641 for (i=0; i<imax; ++i) { 642 ierr = ISDestroy(&is[i]);CHKERRQ(ierr); 643 } 644 645 /* Do Local work */ 646 #if defined(PETSC_USE_CTABLE) 647 ierr = MatIncreaseOverlap_MPIAIJ_Local(C,imax,table,isz,NULL,table_data);CHKERRQ(ierr); 648 #else 649 ierr = MatIncreaseOverlap_MPIAIJ_Local(C,imax,table,isz,data,NULL);CHKERRQ(ierr); 650 #endif 651 652 /* Receive messages */ 653 ierr = PetscMalloc1(nrqr+1,&recv_status);CHKERRQ(ierr); 654 if (nrqr) {ierr = MPI_Waitall(nrqr,r_waits1,recv_status);CHKERRQ(ierr);} 655 656 ierr = PetscMalloc1(nrqs+1,&s_status);CHKERRQ(ierr); 657 if (nrqs) {ierr = MPI_Waitall(nrqs,s_waits1,s_status);CHKERRQ(ierr);} 658 659 /* Phase 1 sends are complete - deallocate buffers */ 660 ierr = PetscFree4(outdat,ptr,tmp,ctr);CHKERRQ(ierr); 661 ierr = PetscFree4(w1,w2,w3,w4);CHKERRQ(ierr); 662 663 ierr = PetscMalloc1(nrqr+1,&xdata);CHKERRQ(ierr); 664 ierr = PetscMalloc1(nrqr+1,&isz1);CHKERRQ(ierr); 665 ierr = MatIncreaseOverlap_MPIAIJ_Receive(C,nrqr,rbuf,xdata,isz1);CHKERRQ(ierr); 666 ierr = PetscFree(rbuf[0]);CHKERRQ(ierr); 667 ierr = PetscFree(rbuf);CHKERRQ(ierr); 668 669 670 /* Send the data back */ 671 /* Do a global reduction to know the buffer space req for incoming messages */ 672 { 673 PetscMPIInt *rw1; 674 675 ierr = PetscCalloc1(size,&rw1);CHKERRQ(ierr); 676 677 for (i=0; i<nrqr; ++i) { 678 proc = recv_status[i].MPI_SOURCE; 679 680 if (proc != onodes1[i]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"MPI_SOURCE mismatch"); 681 rw1[proc] = isz1[i]; 682 } 683 ierr = PetscFree(onodes1);CHKERRQ(ierr); 684 ierr = PetscFree(olengths1);CHKERRQ(ierr); 685 686 /* Determine the number of messages to expect, their lengths, from from-ids */ 687 ierr = PetscGatherMessageLengths(comm,nrqr,nrqs,rw1,&onodes2,&olengths2);CHKERRQ(ierr); 688 ierr = PetscFree(rw1);CHKERRQ(ierr); 689 } 690 /* Now post the Irecvs corresponding to these messages */ 691 ierr = PetscPostIrecvInt(comm,tag2,nrqs,onodes2,olengths2,&rbuf2,&r_waits2);CHKERRQ(ierr); 692 693 /* Now post the sends */ 694 ierr = PetscMalloc1(nrqr+1,&s_waits2);CHKERRQ(ierr); 695 for (i=0; i<nrqr; ++i) { 696 j = recv_status[i].MPI_SOURCE; 697 ierr = MPI_Isend(xdata[i],isz1[i],MPIU_INT,j,tag2,comm,s_waits2+i);CHKERRQ(ierr); 698 } 699 700 /* receive work done on other processors */ 701 { 702 PetscInt is_no,ct1,max,*rbuf2_i,isz_i,jmax; 703 PetscMPIInt idex; 704 PetscBT table_i; 705 MPI_Status *status2; 706 707 ierr = PetscMalloc1((PetscMax(nrqr,nrqs)+1),&status2);CHKERRQ(ierr); 708 for (i=0; i<nrqs; ++i) { 709 ierr = MPI_Waitany(nrqs,r_waits2,&idex,status2+i);CHKERRQ(ierr); 710 /* Process the message */ 711 rbuf2_i = rbuf2[idex]; 712 ct1 = 2*rbuf2_i[0]+1; 713 jmax = rbuf2[idex][0]; 714 for (j=1; j<=jmax; j++) { 715 max = rbuf2_i[2*j]; 716 is_no = rbuf2_i[2*j-1]; 717 isz_i = isz[is_no]; 718 table_i = table[is_no]; 719 #if defined(PETSC_USE_CTABLE) 720 table_data_i = table_data[is_no]; 721 #else 722 data_i = data[is_no]; 723 #endif 724 for (k=0; k<max; k++,ct1++) { 725 row = rbuf2_i[ct1]; 726 if (!PetscBTLookupSet(table_i,row)) { 727 #if defined(PETSC_USE_CTABLE) 728 ierr = PetscTableAdd(table_data_i,row+1,isz_i+1,INSERT_VALUES);CHKERRQ(ierr); 729 #else 730 data_i[isz_i] = row; 731 #endif 732 isz_i++; 733 } 734 } 735 isz[is_no] = isz_i; 736 } 737 } 738 739 if (nrqr) {ierr = MPI_Waitall(nrqr,s_waits2,status2);CHKERRQ(ierr);} 740 ierr = PetscFree(status2);CHKERRQ(ierr); 741 } 742 743 #if defined(PETSC_USE_CTABLE) 744 tcount_max = 0; 745 for (i=0; i<imax; ++i) { 746 table_data_i = table_data[i]; 747 ierr = PetscTableGetCount(table_data_i,&tcount);CHKERRQ(ierr); 748 if (tcount_max < tcount) tcount_max = tcount; 749 } 750 ierr = PetscMalloc1(tcount_max+1,&tdata);CHKERRQ(ierr); 751 #endif 752 753 for (i=0; i<imax; ++i) { 754 #if defined(PETSC_USE_CTABLE) 755 PetscTablePosition tpos; 756 table_data_i = table_data[i]; 757 758 ierr = PetscTableGetHeadPosition(table_data_i,&tpos);CHKERRQ(ierr); 759 while (tpos) { 760 ierr = PetscTableGetNext(table_data_i,&tpos,&k,&j);CHKERRQ(ierr); 761 tdata[--j] = --k; 762 } 763 ierr = ISCreateGeneral(PETSC_COMM_SELF,isz[i],tdata,PETSC_COPY_VALUES,is+i);CHKERRQ(ierr); 764 #else 765 ierr = ISCreateGeneral(PETSC_COMM_SELF,isz[i],data[i],PETSC_COPY_VALUES,is+i);CHKERRQ(ierr); 766 #endif 767 } 768 769 ierr = PetscFree(onodes2);CHKERRQ(ierr); 770 ierr = PetscFree(olengths2);CHKERRQ(ierr); 771 772 ierr = PetscFree(pa);CHKERRQ(ierr); 773 ierr = PetscFree(rbuf2[0]);CHKERRQ(ierr); 774 ierr = PetscFree(rbuf2);CHKERRQ(ierr); 775 ierr = PetscFree(s_waits1);CHKERRQ(ierr); 776 ierr = PetscFree(r_waits1);CHKERRQ(ierr); 777 ierr = PetscFree(s_waits2);CHKERRQ(ierr); 778 ierr = PetscFree(r_waits2);CHKERRQ(ierr); 779 ierr = PetscFree(s_status);CHKERRQ(ierr); 780 ierr = PetscFree(recv_status);CHKERRQ(ierr); 781 ierr = PetscFree(xdata[0]);CHKERRQ(ierr); 782 ierr = PetscFree(xdata);CHKERRQ(ierr); 783 ierr = PetscFree(isz1);CHKERRQ(ierr); 784 #if defined(PETSC_USE_CTABLE) 785 for (i=0; i<imax; i++) { 786 ierr = PetscTableDestroy((PetscTable*)&table_data[i]);CHKERRQ(ierr); 787 } 788 ierr = PetscFree(table_data);CHKERRQ(ierr); 789 ierr = PetscFree(tdata);CHKERRQ(ierr); 790 ierr = PetscFree4(table,data,isz,t_p);CHKERRQ(ierr); 791 #else 792 ierr = PetscFree5(table,data,isz,d_p,t_p);CHKERRQ(ierr); 793 #endif 794 PetscFunctionReturn(0); 795 } 796 797 /* 798 MatIncreaseOverlap_MPIAIJ_Local - Called by MatincreaseOverlap, to do 799 the work on the local processor. 800 801 Inputs: 802 C - MAT_MPIAIJ; 803 imax - total no of index sets processed at a time; 804 table - an array of char - size = m bits. 805 806 Output: 807 isz - array containing the count of the solution elements corresponding 808 to each index set; 809 data or table_data - pointer to the solutions 810 */ 811 static PetscErrorCode MatIncreaseOverlap_MPIAIJ_Local(Mat C,PetscInt imax,PetscBT *table,PetscInt *isz,PetscInt **data,PetscTable *table_data) 812 { 813 Mat_MPIAIJ *c = (Mat_MPIAIJ*)C->data; 814 Mat A = c->A,B = c->B; 815 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data,*b = (Mat_SeqAIJ*)B->data; 816 PetscInt start,end,val,max,rstart,cstart,*ai,*aj; 817 PetscInt *bi,*bj,*garray,i,j,k,row,isz_i; 818 PetscBT table_i; 819 #if defined(PETSC_USE_CTABLE) 820 PetscTable table_data_i; 821 PetscErrorCode ierr; 822 PetscTablePosition tpos; 823 PetscInt tcount,*tdata; 824 #else 825 PetscInt *data_i; 826 #endif 827 828 PetscFunctionBegin; 829 rstart = C->rmap->rstart; 830 cstart = C->cmap->rstart; 831 ai = a->i; 832 aj = a->j; 833 bi = b->i; 834 bj = b->j; 835 garray = c->garray; 836 837 for (i=0; i<imax; i++) { 838 #if defined(PETSC_USE_CTABLE) 839 /* copy existing entries of table_data_i into tdata[] */ 840 table_data_i = table_data[i]; 841 ierr = PetscTableGetCount(table_data_i,&tcount);CHKERRQ(ierr); 842 if (tcount != isz[i]) SETERRQ3(PETSC_COMM_SELF,0," tcount %d != isz[%d] %d",tcount,i,isz[i]); 843 844 ierr = PetscMalloc1(tcount,&tdata);CHKERRQ(ierr); 845 ierr = PetscTableGetHeadPosition(table_data_i,&tpos);CHKERRQ(ierr); 846 while (tpos) { 847 ierr = PetscTableGetNext(table_data_i,&tpos,&row,&j);CHKERRQ(ierr); 848 tdata[--j] = --row; 849 if (j > tcount - 1) SETERRQ2(PETSC_COMM_SELF,0," j %d >= tcount %d",j,tcount); 850 } 851 #else 852 data_i = data[i]; 853 #endif 854 table_i = table[i]; 855 isz_i = isz[i]; 856 max = isz[i]; 857 858 for (j=0; j<max; j++) { 859 #if defined(PETSC_USE_CTABLE) 860 row = tdata[j] - rstart; 861 #else 862 row = data_i[j] - rstart; 863 #endif 864 start = ai[row]; 865 end = ai[row+1]; 866 for (k=start; k<end; k++) { /* Amat */ 867 val = aj[k] + cstart; 868 if (!PetscBTLookupSet(table_i,val)) { 869 #if defined(PETSC_USE_CTABLE) 870 ierr = PetscTableAdd(table_data_i,val+1,isz_i+1,INSERT_VALUES);CHKERRQ(ierr); 871 #else 872 data_i[isz_i] = val; 873 #endif 874 isz_i++; 875 } 876 } 877 start = bi[row]; 878 end = bi[row+1]; 879 for (k=start; k<end; k++) { /* Bmat */ 880 val = garray[bj[k]]; 881 if (!PetscBTLookupSet(table_i,val)) { 882 #if defined(PETSC_USE_CTABLE) 883 ierr = PetscTableAdd(table_data_i,val+1,isz_i+1,INSERT_VALUES);CHKERRQ(ierr); 884 #else 885 data_i[isz_i] = val; 886 #endif 887 isz_i++; 888 } 889 } 890 } 891 isz[i] = isz_i; 892 893 #if defined(PETSC_USE_CTABLE) 894 ierr = PetscFree(tdata);CHKERRQ(ierr); 895 #endif 896 } 897 PetscFunctionReturn(0); 898 } 899 900 /* 901 MatIncreaseOverlap_MPIAIJ_Receive - Process the recieved messages, 902 and return the output 903 904 Input: 905 C - the matrix 906 nrqr - no of messages being processed. 907 rbuf - an array of pointers to the recieved requests 908 909 Output: 910 xdata - array of messages to be sent back 911 isz1 - size of each message 912 913 For better efficiency perhaps we should malloc separately each xdata[i], 914 then if a remalloc is required we need only copy the data for that one row 915 rather then all previous rows as it is now where a single large chunck of 916 memory is used. 917 918 */ 919 static PetscErrorCode MatIncreaseOverlap_MPIAIJ_Receive(Mat C,PetscInt nrqr,PetscInt **rbuf,PetscInt **xdata,PetscInt * isz1) 920 { 921 Mat_MPIAIJ *c = (Mat_MPIAIJ*)C->data; 922 Mat A = c->A,B = c->B; 923 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data,*b = (Mat_SeqAIJ*)B->data; 924 PetscErrorCode ierr; 925 PetscInt rstart,cstart,*ai,*aj,*bi,*bj,*garray,i,j,k; 926 PetscInt row,total_sz,ct,ct1,ct2,ct3,mem_estimate,oct2,l,start,end; 927 PetscInt val,max1,max2,m,no_malloc =0,*tmp,new_estimate,ctr; 928 PetscInt *rbuf_i,kmax,rbuf_0; 929 PetscBT xtable; 930 931 PetscFunctionBegin; 932 m = C->rmap->N; 933 rstart = C->rmap->rstart; 934 cstart = C->cmap->rstart; 935 ai = a->i; 936 aj = a->j; 937 bi = b->i; 938 bj = b->j; 939 garray = c->garray; 940 941 942 for (i=0,ct=0,total_sz=0; i<nrqr; ++i) { 943 rbuf_i = rbuf[i]; 944 rbuf_0 = rbuf_i[0]; 945 ct += rbuf_0; 946 for (j=1; j<=rbuf_0; j++) total_sz += rbuf_i[2*j]; 947 } 948 949 if (C->rmap->n) max1 = ct*(a->nz + b->nz)/C->rmap->n; 950 else max1 = 1; 951 mem_estimate = 3*((total_sz > max1 ? total_sz : max1)+1); 952 ierr = PetscMalloc1(mem_estimate,&xdata[0]);CHKERRQ(ierr); 953 ++no_malloc; 954 ierr = PetscBTCreate(m,&xtable);CHKERRQ(ierr); 955 ierr = PetscMemzero(isz1,nrqr*sizeof(PetscInt));CHKERRQ(ierr); 956 957 ct3 = 0; 958 for (i=0; i<nrqr; i++) { /* for easch mesg from proc i */ 959 rbuf_i = rbuf[i]; 960 rbuf_0 = rbuf_i[0]; 961 ct1 = 2*rbuf_0+1; 962 ct2 = ct1; 963 ct3 += ct1; 964 for (j=1; j<=rbuf_0; j++) { /* for each IS from proc i*/ 965 ierr = PetscBTMemzero(m,xtable);CHKERRQ(ierr); 966 oct2 = ct2; 967 kmax = rbuf_i[2*j]; 968 for (k=0; k<kmax; k++,ct1++) { 969 row = rbuf_i[ct1]; 970 if (!PetscBTLookupSet(xtable,row)) { 971 if (!(ct3 < mem_estimate)) { 972 new_estimate = (PetscInt)(1.5*mem_estimate)+1; 973 ierr = PetscMalloc1(new_estimate,&tmp);CHKERRQ(ierr); 974 ierr = PetscMemcpy(tmp,xdata[0],mem_estimate*sizeof(PetscInt));CHKERRQ(ierr); 975 ierr = PetscFree(xdata[0]);CHKERRQ(ierr); 976 xdata[0] = tmp; 977 mem_estimate = new_estimate; ++no_malloc; 978 for (ctr=1; ctr<=i; ctr++) xdata[ctr] = xdata[ctr-1] + isz1[ctr-1]; 979 } 980 xdata[i][ct2++] = row; 981 ct3++; 982 } 983 } 984 for (k=oct2,max2=ct2; k<max2; k++) { 985 row = xdata[i][k] - rstart; 986 start = ai[row]; 987 end = ai[row+1]; 988 for (l=start; l<end; l++) { 989 val = aj[l] + cstart; 990 if (!PetscBTLookupSet(xtable,val)) { 991 if (!(ct3 < mem_estimate)) { 992 new_estimate = (PetscInt)(1.5*mem_estimate)+1; 993 ierr = PetscMalloc1(new_estimate,&tmp);CHKERRQ(ierr); 994 ierr = PetscMemcpy(tmp,xdata[0],mem_estimate*sizeof(PetscInt));CHKERRQ(ierr); 995 ierr = PetscFree(xdata[0]);CHKERRQ(ierr); 996 xdata[0] = tmp; 997 mem_estimate = new_estimate; ++no_malloc; 998 for (ctr=1; ctr<=i; ctr++) xdata[ctr] = xdata[ctr-1] + isz1[ctr-1]; 999 } 1000 xdata[i][ct2++] = val; 1001 ct3++; 1002 } 1003 } 1004 start = bi[row]; 1005 end = bi[row+1]; 1006 for (l=start; l<end; l++) { 1007 val = garray[bj[l]]; 1008 if (!PetscBTLookupSet(xtable,val)) { 1009 if (!(ct3 < mem_estimate)) { 1010 new_estimate = (PetscInt)(1.5*mem_estimate)+1; 1011 ierr = PetscMalloc1(new_estimate,&tmp);CHKERRQ(ierr); 1012 ierr = PetscMemcpy(tmp,xdata[0],mem_estimate*sizeof(PetscInt));CHKERRQ(ierr); 1013 ierr = PetscFree(xdata[0]);CHKERRQ(ierr); 1014 xdata[0] = tmp; 1015 mem_estimate = new_estimate; ++no_malloc; 1016 for (ctr =1; ctr <=i; ctr++) xdata[ctr] = xdata[ctr-1] + isz1[ctr-1]; 1017 } 1018 xdata[i][ct2++] = val; 1019 ct3++; 1020 } 1021 } 1022 } 1023 /* Update the header*/ 1024 xdata[i][2*j] = ct2 - oct2; /* Undo the vector isz1 and use only a var*/ 1025 xdata[i][2*j-1] = rbuf_i[2*j-1]; 1026 } 1027 xdata[i][0] = rbuf_0; 1028 xdata[i+1] = xdata[i] + ct2; 1029 isz1[i] = ct2; /* size of each message */ 1030 } 1031 ierr = PetscBTDestroy(&xtable);CHKERRQ(ierr); 1032 ierr = PetscInfo3(C,"Allocated %D bytes, required %D bytes, no of mallocs = %D\n",mem_estimate,ct3,no_malloc);CHKERRQ(ierr); 1033 PetscFunctionReturn(0); 1034 } 1035 /* -------------------------------------------------------------------------*/ 1036 extern PetscErrorCode MatGetSubMatrices_MPIAIJ_Local(Mat,PetscInt,const IS[],const IS[],MatReuse,PetscBool*,Mat*); 1037 extern PetscErrorCode MatAssemblyEnd_SeqAIJ(Mat,MatAssemblyType); 1038 /* 1039 Every processor gets the entire matrix 1040 */ 1041 PetscErrorCode MatGetSubMatrix_MPIAIJ_All(Mat A,MatGetSubMatrixOption flag,MatReuse scall,Mat *Bin[]) 1042 { 1043 Mat B; 1044 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 1045 Mat_SeqAIJ *b,*ad = (Mat_SeqAIJ*)a->A->data,*bd = (Mat_SeqAIJ*)a->B->data; 1046 PetscErrorCode ierr; 1047 PetscMPIInt size,rank,*recvcounts = 0,*displs = 0; 1048 PetscInt sendcount,i,*rstarts = A->rmap->range,n,cnt,j; 1049 PetscInt m,*b_sendj,*garray = a->garray,*lens,*jsendbuf,*a_jsendbuf,*b_jsendbuf; 1050 MatScalar *sendbuf,*recvbuf,*a_sendbuf,*b_sendbuf; 1051 1052 PetscFunctionBegin; 1053 ierr = MPI_Comm_size(PetscObjectComm((PetscObject)A),&size);CHKERRQ(ierr); 1054 ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)A),&rank);CHKERRQ(ierr); 1055 1056 if (scall == MAT_INITIAL_MATRIX) { 1057 /* ---------------------------------------------------------------- 1058 Tell every processor the number of nonzeros per row 1059 */ 1060 ierr = PetscMalloc1(A->rmap->N,&lens);CHKERRQ(ierr); 1061 for (i=A->rmap->rstart; i<A->rmap->rend; i++) { 1062 lens[i] = ad->i[i-A->rmap->rstart+1] - ad->i[i-A->rmap->rstart] + bd->i[i-A->rmap->rstart+1] - bd->i[i-A->rmap->rstart]; 1063 } 1064 ierr = PetscMalloc2(size,&recvcounts,size,&displs);CHKERRQ(ierr); 1065 for (i=0; i<size; i++) { 1066 recvcounts[i] = A->rmap->range[i+1] - A->rmap->range[i]; 1067 displs[i] = A->rmap->range[i]; 1068 } 1069 #if defined(PETSC_HAVE_MPI_IN_PLACE) 1070 ierr = MPI_Allgatherv(MPI_IN_PLACE,0,MPI_DATATYPE_NULL,lens,recvcounts,displs,MPIU_INT,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 1071 #else 1072 sendcount = A->rmap->rend - A->rmap->rstart; 1073 ierr = MPI_Allgatherv(lens+A->rmap->rstart,sendcount,MPIU_INT,lens,recvcounts,displs,MPIU_INT,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 1074 #endif 1075 /* --------------------------------------------------------------- 1076 Create the sequential matrix of the same type as the local block diagonal 1077 */ 1078 ierr = MatCreate(PETSC_COMM_SELF,&B);CHKERRQ(ierr); 1079 ierr = MatSetSizes(B,A->rmap->N,A->cmap->N,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 1080 ierr = MatSetBlockSizesFromMats(B,A,A);CHKERRQ(ierr); 1081 ierr = MatSetType(B,((PetscObject)a->A)->type_name);CHKERRQ(ierr); 1082 ierr = MatSeqAIJSetPreallocation(B,0,lens);CHKERRQ(ierr); 1083 ierr = PetscMalloc1(1,Bin);CHKERRQ(ierr); 1084 **Bin = B; 1085 b = (Mat_SeqAIJ*)B->data; 1086 1087 /*-------------------------------------------------------------------- 1088 Copy my part of matrix column indices over 1089 */ 1090 sendcount = ad->nz + bd->nz; 1091 jsendbuf = b->j + b->i[rstarts[rank]]; 1092 a_jsendbuf = ad->j; 1093 b_jsendbuf = bd->j; 1094 n = A->rmap->rend - A->rmap->rstart; 1095 cnt = 0; 1096 for (i=0; i<n; i++) { 1097 1098 /* put in lower diagonal portion */ 1099 m = bd->i[i+1] - bd->i[i]; 1100 while (m > 0) { 1101 /* is it above diagonal (in bd (compressed) numbering) */ 1102 if (garray[*b_jsendbuf] > A->rmap->rstart + i) break; 1103 jsendbuf[cnt++] = garray[*b_jsendbuf++]; 1104 m--; 1105 } 1106 1107 /* put in diagonal portion */ 1108 for (j=ad->i[i]; j<ad->i[i+1]; j++) { 1109 jsendbuf[cnt++] = A->rmap->rstart + *a_jsendbuf++; 1110 } 1111 1112 /* put in upper diagonal portion */ 1113 while (m-- > 0) { 1114 jsendbuf[cnt++] = garray[*b_jsendbuf++]; 1115 } 1116 } 1117 if (cnt != sendcount) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Corrupted PETSc matrix: nz given %D actual nz %D",sendcount,cnt); 1118 1119 /*-------------------------------------------------------------------- 1120 Gather all column indices to all processors 1121 */ 1122 for (i=0; i<size; i++) { 1123 recvcounts[i] = 0; 1124 for (j=A->rmap->range[i]; j<A->rmap->range[i+1]; j++) { 1125 recvcounts[i] += lens[j]; 1126 } 1127 } 1128 displs[0] = 0; 1129 for (i=1; i<size; i++) { 1130 displs[i] = displs[i-1] + recvcounts[i-1]; 1131 } 1132 #if defined(PETSC_HAVE_MPI_IN_PLACE) 1133 ierr = MPI_Allgatherv(MPI_IN_PLACE,0,MPI_DATATYPE_NULL,b->j,recvcounts,displs,MPIU_INT,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 1134 #else 1135 ierr = MPI_Allgatherv(jsendbuf,sendcount,MPIU_INT,b->j,recvcounts,displs,MPIU_INT,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 1136 #endif 1137 /*-------------------------------------------------------------------- 1138 Assemble the matrix into useable form (note numerical values not yet set) 1139 */ 1140 /* set the b->ilen (length of each row) values */ 1141 ierr = PetscMemcpy(b->ilen,lens,A->rmap->N*sizeof(PetscInt));CHKERRQ(ierr); 1142 /* set the b->i indices */ 1143 b->i[0] = 0; 1144 for (i=1; i<=A->rmap->N; i++) { 1145 b->i[i] = b->i[i-1] + lens[i-1]; 1146 } 1147 ierr = PetscFree(lens);CHKERRQ(ierr); 1148 ierr = MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1149 ierr = MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1150 1151 } else { 1152 B = **Bin; 1153 b = (Mat_SeqAIJ*)B->data; 1154 } 1155 1156 /*-------------------------------------------------------------------- 1157 Copy my part of matrix numerical values into the values location 1158 */ 1159 if (flag == MAT_GET_VALUES) { 1160 sendcount = ad->nz + bd->nz; 1161 sendbuf = b->a + b->i[rstarts[rank]]; 1162 a_sendbuf = ad->a; 1163 b_sendbuf = bd->a; 1164 b_sendj = bd->j; 1165 n = A->rmap->rend - A->rmap->rstart; 1166 cnt = 0; 1167 for (i=0; i<n; i++) { 1168 1169 /* put in lower diagonal portion */ 1170 m = bd->i[i+1] - bd->i[i]; 1171 while (m > 0) { 1172 /* is it above diagonal (in bd (compressed) numbering) */ 1173 if (garray[*b_sendj] > A->rmap->rstart + i) break; 1174 sendbuf[cnt++] = *b_sendbuf++; 1175 m--; 1176 b_sendj++; 1177 } 1178 1179 /* put in diagonal portion */ 1180 for (j=ad->i[i]; j<ad->i[i+1]; j++) { 1181 sendbuf[cnt++] = *a_sendbuf++; 1182 } 1183 1184 /* put in upper diagonal portion */ 1185 while (m-- > 0) { 1186 sendbuf[cnt++] = *b_sendbuf++; 1187 b_sendj++; 1188 } 1189 } 1190 if (cnt != sendcount) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Corrupted PETSc matrix: nz given %D actual nz %D",sendcount,cnt); 1191 1192 /* ----------------------------------------------------------------- 1193 Gather all numerical values to all processors 1194 */ 1195 if (!recvcounts) { 1196 ierr = PetscMalloc2(size,&recvcounts,size,&displs);CHKERRQ(ierr); 1197 } 1198 for (i=0; i<size; i++) { 1199 recvcounts[i] = b->i[rstarts[i+1]] - b->i[rstarts[i]]; 1200 } 1201 displs[0] = 0; 1202 for (i=1; i<size; i++) { 1203 displs[i] = displs[i-1] + recvcounts[i-1]; 1204 } 1205 recvbuf = b->a; 1206 #if defined(PETSC_HAVE_MPI_IN_PLACE) 1207 ierr = MPI_Allgatherv(MPI_IN_PLACE,0,MPI_DATATYPE_NULL,recvbuf,recvcounts,displs,MPIU_SCALAR,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 1208 #else 1209 ierr = MPI_Allgatherv(sendbuf,sendcount,MPIU_SCALAR,recvbuf,recvcounts,displs,MPIU_SCALAR,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 1210 #endif 1211 } /* endof (flag == MAT_GET_VALUES) */ 1212 ierr = PetscFree2(recvcounts,displs);CHKERRQ(ierr); 1213 1214 if (A->symmetric) { 1215 ierr = MatSetOption(B,MAT_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr); 1216 } else if (A->hermitian) { 1217 ierr = MatSetOption(B,MAT_HERMITIAN,PETSC_TRUE);CHKERRQ(ierr); 1218 } else if (A->structurally_symmetric) { 1219 ierr = MatSetOption(B,MAT_STRUCTURALLY_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr); 1220 } 1221 PetscFunctionReturn(0); 1222 } 1223 1224 PetscErrorCode MatDestroy_MPIAIJ_MatGetSubmatrices(Mat C) 1225 { 1226 PetscErrorCode ierr; 1227 Mat_SeqAIJ *c = (Mat_SeqAIJ*)C->data; 1228 Mat_SubMat *submatj = c->submatis1; 1229 PetscInt i; 1230 1231 PetscFunctionBegin; 1232 if (!submatj->id) { /* delete data that are linked only to submats[id=0] */ 1233 ierr = PetscFree4(submatj->sbuf1,submatj->ptr,submatj->tmp,submatj->ctr);CHKERRQ(ierr); 1234 1235 for (i=0; i<submatj->nrqr; ++i) { 1236 ierr = PetscFree(submatj->sbuf2[i]);CHKERRQ(ierr); 1237 } 1238 ierr = PetscFree3(submatj->sbuf2,submatj->req_size,submatj->req_source1);CHKERRQ(ierr); 1239 1240 if (submatj->rbuf1) { 1241 ierr = PetscFree(submatj->rbuf1[0]);CHKERRQ(ierr); 1242 ierr = PetscFree(submatj->rbuf1);CHKERRQ(ierr); 1243 } 1244 1245 for (i=0; i<submatj->nrqs; ++i) { 1246 ierr = PetscFree(submatj->rbuf3[i]);CHKERRQ(ierr); 1247 } 1248 ierr = PetscFree3(submatj->req_source2,submatj->rbuf2,submatj->rbuf3);CHKERRQ(ierr); 1249 ierr = PetscFree(submatj->pa);CHKERRQ(ierr); 1250 } 1251 1252 #if defined(PETSC_USE_CTABLE) 1253 ierr = PetscTableDestroy((PetscTable*)&submatj->rmap);CHKERRQ(ierr); 1254 if (submatj->cmap_loc) {ierr = PetscFree(submatj->cmap_loc);CHKERRQ(ierr);} 1255 ierr = PetscFree(submatj->rmap_loc);CHKERRQ(ierr); 1256 #else 1257 ierr = PetscFree(submatj->rmap);CHKERRQ(ierr); 1258 #endif 1259 1260 if (!submatj->allcolumns) { 1261 #if defined(PETSC_USE_CTABLE) 1262 ierr = PetscTableDestroy((PetscTable*)&submatj->cmap);CHKERRQ(ierr); 1263 #else 1264 ierr = PetscFree(submatj->cmap);CHKERRQ(ierr); 1265 #endif 1266 } 1267 ierr = submatj->destroy(C);CHKERRQ(ierr); 1268 ierr = PetscFree(submatj->row2proc);CHKERRQ(ierr); 1269 1270 ierr = PetscFree(submatj);CHKERRQ(ierr); 1271 PetscFunctionReturn(0); 1272 } 1273 1274 PetscErrorCode MatGetSubMatrices_MPIAIJ_SingleIS_Local(Mat C,PetscInt ismax,const IS isrow[],const IS iscol[],MatReuse scall,PetscBool allcolumns,Mat *submats) 1275 { 1276 Mat_MPIAIJ *c = (Mat_MPIAIJ*)C->data; 1277 Mat submat,A = c->A,B = c->B; 1278 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data,*b = (Mat_SeqAIJ*)B->data,*subc; 1279 PetscInt *ai=a->i,*aj=a->j,*bi=b->i,*bj=b->j,nzA,nzB; 1280 PetscInt cstart = C->cmap->rstart,cend = C->cmap->rend,rstart = C->rmap->rstart,*bmap = c->garray; 1281 const PetscInt *icol,*irow; 1282 PetscInt nrow,ncol,start; 1283 PetscErrorCode ierr; 1284 PetscMPIInt rank,size,tag1,tag2,tag3,tag4,*w1,*w2,nrqr; 1285 PetscInt **sbuf1,**sbuf2,i,j,k,l,ct1,ct2,ct3,**rbuf1,row,proc; 1286 PetscInt nrqs=0,msz,**ptr,*req_size,*ctr,*pa,*tmp,tcol,*iptr; 1287 PetscInt **rbuf3,*req_source1,*req_source2,**sbuf_aj,**rbuf2,max1,nnz; 1288 PetscInt *lens,rmax,ncols,*cols,Crow; 1289 #if defined(PETSC_USE_CTABLE) 1290 PetscTable cmap,rmap; 1291 PetscInt *cmap_loc,*rmap_loc; 1292 #else 1293 PetscInt *cmap,*rmap; 1294 #endif 1295 PetscInt ctr_j,*sbuf1_j,*sbuf_aj_i,*rbuf1_i,kmax,*sbuf1_i,*rbuf2_i,*rbuf3_i; 1296 PetscInt *cworkB,lwrite,*subcols,*row2proc; 1297 PetscScalar *vworkA,*vworkB,*a_a = a->a,*b_a = b->a,*subvals=NULL; 1298 MPI_Request *s_waits1,*r_waits1,*s_waits2,*r_waits2,*r_waits3; 1299 MPI_Request *r_waits4,*s_waits3 = NULL,*s_waits4; 1300 MPI_Status *r_status1,*r_status2,*s_status1,*s_status3 = NULL,*s_status2; 1301 MPI_Status *r_status3 = NULL,*r_status4,*s_status4; 1302 MPI_Comm comm; 1303 PetscScalar **rbuf4,**sbuf_aa,*vals,*sbuf_aa_i,*rbuf4_i; 1304 PetscMPIInt *onodes1,*olengths1,idex,end; 1305 Mat_SubMat *smatis1; 1306 PetscBool isrowsorted; 1307 1308 PetscFunctionBegin; 1309 if (ismax != 1) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"This routine only works when all processes have ismax=1"); 1310 1311 ierr = PetscObjectGetComm((PetscObject)C,&comm);CHKERRQ(ierr); 1312 size = c->size; 1313 rank = c->rank; 1314 1315 ierr = ISSorted(isrow[0],&isrowsorted);CHKERRQ(ierr); 1316 if (!isrowsorted) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"isrow[0] must be sorted"); 1317 1318 ierr = ISGetIndices(isrow[0],&irow);CHKERRQ(ierr); 1319 ierr = ISGetLocalSize(isrow[0],&nrow);CHKERRQ(ierr); 1320 if (allcolumns) { 1321 icol = NULL; 1322 ncol = C->cmap->N; 1323 } else { 1324 ierr = ISGetIndices(iscol[0],&icol);CHKERRQ(ierr); 1325 ierr = ISGetLocalSize(iscol[0],&ncol);CHKERRQ(ierr); 1326 } 1327 1328 if (scall == MAT_INITIAL_MATRIX) { 1329 PetscInt *sbuf2_i,*cworkA,lwrite,ctmp; 1330 1331 /* Get some new tags to keep the communication clean */ 1332 tag1 = ((PetscObject)C)->tag; 1333 ierr = PetscObjectGetNewTag((PetscObject)C,&tag2);CHKERRQ(ierr); 1334 ierr = PetscObjectGetNewTag((PetscObject)C,&tag3);CHKERRQ(ierr); 1335 1336 /* evaluate communication - mesg to who, length of mesg, and buffer space 1337 required. Based on this, buffers are allocated, and data copied into them */ 1338 ierr = PetscCalloc2(size,&w1,size,&w2);CHKERRQ(ierr); 1339 ierr = PetscMalloc1(nrow,&row2proc);CHKERRQ(ierr); 1340 1341 /* w1[proc] = num of rows owned by proc -- to be requested */ 1342 proc = 0; 1343 nrqs = 0; /* num of outgoing messages */ 1344 for (j=0; j<nrow; j++) { 1345 row = irow[j]; /* sorted! */ 1346 while (row >= C->rmap->range[proc+1]) proc++; 1347 w1[proc]++; 1348 row2proc[j] = proc; /* map row index to proc */ 1349 1350 if (proc != rank && !w2[proc]) { 1351 w2[proc] = 1; nrqs++; 1352 } 1353 } 1354 w1[rank] = 0; /* rows owned by self will not be requested */ 1355 1356 ierr = PetscMalloc1(nrqs+1,&pa);CHKERRQ(ierr); /*(proc -array)*/ 1357 for (proc=0,j=0; proc<size; proc++) { 1358 if (w1[proc]) { pa[j++] = proc;} 1359 } 1360 1361 /* Each message would have a header = 1 + 2*(num of IS) + data (here,num of IS = 1) */ 1362 msz = 0; /* total mesg length (for all procs) */ 1363 for (i=0; i<nrqs; i++) { 1364 proc = pa[i]; 1365 w1[proc] += 3; 1366 msz += w1[proc]; 1367 } 1368 ierr = PetscInfo2(0,"Number of outgoing messages %D Total message length %D\n",nrqs,msz);CHKERRQ(ierr); 1369 1370 /* Determine nrqr, the number of messages to expect, their lengths, from from-ids */ 1371 /* if w2[proc]=1, a message of length w1[proc] will be sent to proc; */ 1372 ierr = PetscGatherNumberOfMessages(comm,w2,w1,&nrqr);CHKERRQ(ierr); 1373 1374 /* Input: nrqs: nsend; nrqr: nrecv; w1: msg length to be sent; 1375 Output: onodes1: recv node-ids; olengths1: corresponding recv message length */ 1376 ierr = PetscGatherMessageLengths(comm,nrqs,nrqr,w1,&onodes1,&olengths1);CHKERRQ(ierr); 1377 1378 /* Now post the Irecvs corresponding to these messages */ 1379 ierr = PetscPostIrecvInt(comm,tag1,nrqr,onodes1,olengths1,&rbuf1,&r_waits1);CHKERRQ(ierr); 1380 1381 ierr = PetscFree(onodes1);CHKERRQ(ierr); 1382 ierr = PetscFree(olengths1);CHKERRQ(ierr); 1383 1384 /* Allocate Memory for outgoing messages */ 1385 ierr = PetscMalloc4(size,&sbuf1,size,&ptr,2*msz,&tmp,size,&ctr);CHKERRQ(ierr); 1386 ierr = PetscMemzero(sbuf1,size*sizeof(PetscInt*));CHKERRQ(ierr); 1387 ierr = PetscMemzero(ptr,size*sizeof(PetscInt*));CHKERRQ(ierr); 1388 1389 /* subf1[pa[0]] = tmp, subf1[pa[i]] = subf1[pa[i-1]] + w1[pa[i-1]] */ 1390 iptr = tmp; 1391 for (i=0; i<nrqs; i++) { 1392 proc = pa[i]; 1393 sbuf1[proc] = iptr; 1394 iptr += w1[proc]; 1395 } 1396 1397 /* Form the outgoing messages */ 1398 /* Initialize the header space */ 1399 for (i=0; i<nrqs; i++) { 1400 proc = pa[i]; 1401 ierr = PetscMemzero(sbuf1[proc],3*sizeof(PetscInt));CHKERRQ(ierr); 1402 ptr[proc] = sbuf1[proc] + 3; 1403 } 1404 1405 /* Parse the isrow and copy data into outbuf */ 1406 ierr = PetscMemzero(ctr,size*sizeof(PetscInt));CHKERRQ(ierr); 1407 for (j=0; j<nrow; j++) { /* parse the indices of each IS */ 1408 proc = row2proc[j]; 1409 if (proc != rank) { /* copy to the outgoing buf*/ 1410 *ptr[proc] = irow[j]; 1411 ctr[proc]++; ptr[proc]++; 1412 } 1413 } 1414 1415 /* Update the headers for the current IS */ 1416 for (j=0; j<size; j++) { /* Can Optimise this loop too */ 1417 if ((ctr_j = ctr[j])) { 1418 sbuf1_j = sbuf1[j]; 1419 k = ++sbuf1_j[0]; 1420 sbuf1_j[2*k] = ctr_j; 1421 sbuf1_j[2*k-1] = 0; 1422 } 1423 } 1424 1425 /* Now post the sends */ 1426 ierr = PetscMalloc1(nrqs+1,&s_waits1);CHKERRQ(ierr); 1427 for (i=0; i<nrqs; ++i) { 1428 proc = pa[i]; 1429 ierr = MPI_Isend(sbuf1[proc],w1[proc],MPIU_INT,proc,tag1,comm,s_waits1+i);CHKERRQ(ierr); 1430 } 1431 1432 /* Post Receives to capture the buffer size */ 1433 ierr = PetscMalloc4(nrqs+1,&r_status2,nrqr+1,&s_waits2,nrqs+1,&r_waits2,nrqr+1,&s_status2);CHKERRQ(ierr); 1434 ierr = PetscMalloc3(nrqs+1,&req_source2,nrqs+1,&rbuf2,nrqs+1,&rbuf3);CHKERRQ(ierr); 1435 1436 rbuf2[0] = tmp + msz; 1437 for (i=1; i<nrqs; ++i) rbuf2[i] = rbuf2[i-1] + w1[pa[i-1]]; 1438 1439 for (i=0; i<nrqs; ++i) { 1440 proc = pa[i]; 1441 ierr = MPI_Irecv(rbuf2[i],w1[proc],MPIU_INT,proc,tag2,comm,r_waits2+i);CHKERRQ(ierr); 1442 } 1443 1444 ierr = PetscFree2(w1,w2);CHKERRQ(ierr); 1445 1446 /* Send to other procs the buf size they should allocate */ 1447 /* Receive messages*/ 1448 ierr = PetscMalloc1(nrqr+1,&r_status1);CHKERRQ(ierr); 1449 ierr = PetscMalloc3(nrqr,&sbuf2,nrqr,&req_size,nrqr,&req_source1);CHKERRQ(ierr); 1450 1451 ierr = MPI_Waitall(nrqr,r_waits1,r_status1);CHKERRQ(ierr); 1452 for (i=0; i<nrqr; ++i) { 1453 req_size[i] = 0; 1454 rbuf1_i = rbuf1[i]; 1455 start = 2*rbuf1_i[0] + 1; 1456 ierr = MPI_Get_count(r_status1+i,MPIU_INT,&end);CHKERRQ(ierr); 1457 ierr = PetscMalloc1(end+1,&sbuf2[i]);CHKERRQ(ierr); 1458 sbuf2_i = sbuf2[i]; 1459 for (j=start; j<end; j++) { 1460 k = rbuf1_i[j] - rstart; 1461 ncols = ai[k+1] - ai[k] + bi[k+1] - bi[k]; 1462 sbuf2_i[j] = ncols; 1463 req_size[i] += ncols; 1464 } 1465 req_source1[i] = r_status1[i].MPI_SOURCE; 1466 1467 /* form the header */ 1468 sbuf2_i[0] = req_size[i]; 1469 for (j=1; j<start; j++) sbuf2_i[j] = rbuf1_i[j]; 1470 1471 ierr = MPI_Isend(sbuf2_i,end,MPIU_INT,req_source1[i],tag2,comm,s_waits2+i);CHKERRQ(ierr); 1472 } 1473 1474 ierr = PetscFree(r_status1);CHKERRQ(ierr); 1475 ierr = PetscFree(r_waits1);CHKERRQ(ierr); 1476 1477 /* rbuf2 is received, Post recv column indices a->j */ 1478 ierr = MPI_Waitall(nrqs,r_waits2,r_status2);CHKERRQ(ierr); 1479 1480 ierr = PetscMalloc4(nrqs+1,&r_waits3,nrqr+1,&s_waits3,nrqs+1,&r_status3,nrqr+1,&s_status3);CHKERRQ(ierr); 1481 for (i=0; i<nrqs; ++i) { 1482 ierr = PetscMalloc1(rbuf2[i][0]+1,&rbuf3[i]);CHKERRQ(ierr); 1483 req_source2[i] = r_status2[i].MPI_SOURCE; 1484 ierr = MPI_Irecv(rbuf3[i],rbuf2[i][0],MPIU_INT,req_source2[i],tag3,comm,r_waits3+i);CHKERRQ(ierr); 1485 } 1486 1487 /* Wait on sends1 and sends2 */ 1488 ierr = PetscMalloc1(nrqs+1,&s_status1);CHKERRQ(ierr); 1489 ierr = MPI_Waitall(nrqs,s_waits1,s_status1);CHKERRQ(ierr); 1490 ierr = PetscFree(s_waits1);CHKERRQ(ierr); 1491 ierr = PetscFree(s_status1);CHKERRQ(ierr); 1492 1493 ierr = MPI_Waitall(nrqr,s_waits2,s_status2);CHKERRQ(ierr); 1494 ierr = PetscFree4(r_status2,s_waits2,r_waits2,s_status2);CHKERRQ(ierr); 1495 1496 /* Now allocate sending buffers for a->j, and send them off */ 1497 ierr = PetscMalloc1(nrqr+1,&sbuf_aj);CHKERRQ(ierr); 1498 for (i=0,j=0; i<nrqr; i++) j += req_size[i]; 1499 ierr = PetscMalloc1(j+1,&sbuf_aj[0]);CHKERRQ(ierr); 1500 for (i=1; i<nrqr; i++) sbuf_aj[i] = sbuf_aj[i-1] + req_size[i-1]; 1501 1502 for (i=0; i<nrqr; i++) { /* for each requested message */ 1503 rbuf1_i = rbuf1[i]; 1504 sbuf_aj_i = sbuf_aj[i]; 1505 ct1 = 2*rbuf1_i[0] + 1; 1506 ct2 = 0; 1507 /* max1=rbuf1_i[0]; if (max1 != 1) SETERRQ1(PETSC_COMM_SELF,0,"max1 %d != 1",max1); */ 1508 1509 kmax = rbuf1[i][2]; 1510 for (k=0; k<kmax; k++,ct1++) { /* for each row */ 1511 row = rbuf1_i[ct1] - rstart; 1512 nzA = ai[row+1] - ai[row]; 1513 nzB = bi[row+1] - bi[row]; 1514 ncols = nzA + nzB; 1515 cworkA = aj + ai[row]; cworkB = bj + bi[row]; 1516 1517 /* load the column indices for this row into cols*/ 1518 cols = sbuf_aj_i + ct2; 1519 1520 lwrite = 0; 1521 for (l=0; l<nzB; l++) { 1522 if ((ctmp = bmap[cworkB[l]]) < cstart) cols[lwrite++] = ctmp; 1523 } 1524 for (l=0; l<nzA; l++) cols[lwrite++] = cstart + cworkA[l]; 1525 for (l=0; l<nzB; l++) { 1526 if ((ctmp = bmap[cworkB[l]]) >= cend) cols[lwrite++] = ctmp; 1527 } 1528 1529 ct2 += ncols; 1530 } 1531 ierr = MPI_Isend(sbuf_aj_i,req_size[i],MPIU_INT,req_source1[i],tag3,comm,s_waits3+i);CHKERRQ(ierr); 1532 } 1533 1534 /* create column map (cmap): global col of C -> local col of submat */ 1535 #if defined(PETSC_USE_CTABLE) 1536 if (!allcolumns) { 1537 ierr = PetscTableCreate(ncol+1,C->cmap->N+1,&cmap);CHKERRQ(ierr); 1538 ierr = PetscCalloc1(C->cmap->n,&cmap_loc);CHKERRQ(ierr); 1539 for (j=0; j<ncol; j++) { /* use array cmap_loc[] for local col indices */ 1540 if (icol[j] >= cstart && icol[j] <cend) { 1541 cmap_loc[icol[j] - cstart] = j+1; 1542 } else { /* use PetscTable for non-local col indices */ 1543 ierr = PetscTableAdd(cmap,icol[j]+1,j+1,INSERT_VALUES);CHKERRQ(ierr); 1544 } 1545 } 1546 } else { 1547 cmap = NULL; 1548 cmap_loc = NULL; 1549 } 1550 ierr = PetscCalloc1(C->rmap->n,&rmap_loc);CHKERRQ(ierr); 1551 #else 1552 if (!allcolumns) { 1553 ierr = PetscCalloc1(C->cmap->N,&cmap);CHKERRQ(ierr); 1554 for (j=0; j<ncol; j++) cmap[icol[j]] = j+1; 1555 } else { 1556 cmap = NULL; 1557 } 1558 #endif 1559 1560 /* Create lens for MatSeqAIJSetPreallocation() */ 1561 ierr = PetscCalloc1(nrow,&lens);CHKERRQ(ierr); 1562 1563 /* Compute lens from local part of C */ 1564 for (j=0; j<nrow; j++) { 1565 row = irow[j]; 1566 proc = row2proc[j]; 1567 if (proc == rank) { 1568 /* diagonal part A = c->A */ 1569 ncols = ai[row-rstart+1] - ai[row-rstart]; 1570 cols = aj + ai[row-rstart]; 1571 if (!allcolumns) { 1572 for (k=0; k<ncols; k++) { 1573 #if defined(PETSC_USE_CTABLE) 1574 tcol = cmap_loc[cols[k]]; 1575 #else 1576 tcol = cmap[cols[k]+cstart]; 1577 #endif 1578 if (tcol) lens[j]++; 1579 } 1580 } else { /* allcolumns */ 1581 lens[j] = ncols; 1582 } 1583 1584 /* off-diagonal part B = c->B */ 1585 ncols = bi[row-rstart+1] - bi[row-rstart]; 1586 cols = bj + bi[row-rstart]; 1587 if (!allcolumns) { 1588 for (k=0; k<ncols; k++) { 1589 #if defined(PETSC_USE_CTABLE) 1590 ierr = PetscTableFind(cmap,bmap[cols[k]]+1,&tcol);CHKERRQ(ierr); 1591 #else 1592 tcol = cmap[bmap[cols[k]]]; 1593 #endif 1594 if (tcol) lens[j]++; 1595 } 1596 } else { /* allcolumns */ 1597 lens[j] += ncols; 1598 } 1599 } 1600 } 1601 1602 /* Create row map (rmap): global row of C -> local row of submat */ 1603 #if defined(PETSC_USE_CTABLE) 1604 ierr = PetscTableCreate(nrow+1,C->rmap->N+1,&rmap);CHKERRQ(ierr); 1605 for (j=0; j<nrow; j++) { 1606 row = irow[j]; 1607 proc = row2proc[j]; 1608 if (proc == rank) { /* a local row */ 1609 rmap_loc[row - rstart] = j; 1610 } else { 1611 ierr = PetscTableAdd(rmap,irow[j]+1,j+1,INSERT_VALUES);CHKERRQ(ierr); 1612 } 1613 } 1614 #else 1615 ierr = PetscCalloc1(C->rmap->N,&rmap);CHKERRQ(ierr); 1616 for (j=0; j<nrow; j++) { 1617 rmap[irow[j]] = j; 1618 } 1619 #endif 1620 1621 /* Update lens from offproc data */ 1622 /* recv a->j is done */ 1623 ierr = MPI_Waitall(nrqs,r_waits3,r_status3);CHKERRQ(ierr); 1624 for (i=0; i<nrqs; i++) { 1625 proc = pa[i]; 1626 sbuf1_i = sbuf1[proc]; 1627 /* jmax = sbuf1_i[0]; if (jmax != 1) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"jmax !=1"); */ 1628 ct1 = 2 + 1; 1629 ct2 = 0; 1630 rbuf2_i = rbuf2[i]; /* received length of C->j */ 1631 rbuf3_i = rbuf3[i]; /* received C->j */ 1632 1633 /* is_no = sbuf1_i[2*j-1]; if (is_no != 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"is_no !=0"); */ 1634 max1 = sbuf1_i[2]; 1635 for (k=0; k<max1; k++,ct1++) { 1636 #if defined(PETSC_USE_CTABLE) 1637 ierr = PetscTableFind(rmap,sbuf1_i[ct1]+1,&row);CHKERRQ(ierr); 1638 row--; 1639 if (row < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"row not found in table"); 1640 #else 1641 row = rmap[sbuf1_i[ct1]]; /* the row index in submat */ 1642 #endif 1643 /* Now, store row index of submat in sbuf1_i[ct1] */ 1644 sbuf1_i[ct1] = row; 1645 1646 nnz = rbuf2_i[ct1]; 1647 if (!allcolumns) { 1648 for (l=0; l<nnz; l++,ct2++) { 1649 #if defined(PETSC_USE_CTABLE) 1650 if (rbuf3_i[ct2] >= cstart && rbuf3_i[ct2] <cend) { 1651 tcol = cmap_loc[rbuf3_i[ct2] - cstart]; 1652 } else { 1653 ierr = PetscTableFind(cmap,rbuf3_i[ct2]+1,&tcol);CHKERRQ(ierr); 1654 } 1655 #else 1656 tcol = cmap[rbuf3_i[ct2]]; /* column index in submat */ 1657 #endif 1658 if (tcol) lens[row]++; 1659 } 1660 } else { /* allcolumns */ 1661 lens[row] += nnz; 1662 } 1663 } 1664 } 1665 ierr = MPI_Waitall(nrqr,s_waits3,s_status3);CHKERRQ(ierr); 1666 ierr = PetscFree4(r_waits3,s_waits3,r_status3,s_status3);CHKERRQ(ierr); 1667 1668 /* Create the submatrices */ 1669 ierr = MatCreate(PETSC_COMM_SELF,&submat);CHKERRQ(ierr); 1670 ierr = MatSetSizes(submat,nrow,ncol,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 1671 1672 ierr = ISGetBlockSize(isrow[0],&i);CHKERRQ(ierr); 1673 ierr = ISGetBlockSize(iscol[0],&j);CHKERRQ(ierr); 1674 ierr = MatSetBlockSizes(submat,i,j);CHKERRQ(ierr); 1675 ierr = MatSetType(submat,((PetscObject)A)->type_name);CHKERRQ(ierr); 1676 ierr = MatSeqAIJSetPreallocation(submat,0,lens);CHKERRQ(ierr); 1677 1678 /* create struct Mat_SubMat and attached it to submat */ 1679 ierr = PetscNew(&smatis1);CHKERRQ(ierr); 1680 subc = (Mat_SeqAIJ*)submat->data; 1681 subc->submatis1 = smatis1; 1682 1683 smatis1->id = 0; 1684 smatis1->nrqs = nrqs; 1685 smatis1->nrqr = nrqr; 1686 smatis1->rbuf1 = rbuf1; 1687 smatis1->rbuf2 = rbuf2; 1688 smatis1->rbuf3 = rbuf3; 1689 smatis1->sbuf2 = sbuf2; 1690 smatis1->req_source2 = req_source2; 1691 1692 smatis1->sbuf1 = sbuf1; 1693 smatis1->ptr = ptr; 1694 smatis1->tmp = tmp; 1695 smatis1->ctr = ctr; 1696 1697 smatis1->pa = pa; 1698 smatis1->req_size = req_size; 1699 smatis1->req_source1 = req_source1; 1700 1701 smatis1->allcolumns = allcolumns; 1702 smatis1->row2proc = row2proc; 1703 smatis1->rmap = rmap; 1704 smatis1->cmap = cmap; 1705 #if defined(PETSC_USE_CTABLE) 1706 smatis1->rmap_loc = rmap_loc; 1707 smatis1->cmap_loc = cmap_loc; 1708 #endif 1709 1710 smatis1->destroy = submat->ops->destroy; 1711 submat->ops->destroy = MatDestroy_MPIAIJ_MatGetSubmatrices; 1712 submat->factortype = C->factortype; 1713 1714 /* compute rmax */ 1715 rmax = 0; 1716 for (i=0; i<nrow; i++) rmax = PetscMax(rmax,lens[i]); 1717 1718 } else { /* scall == MAT_REUSE_MATRIX */ 1719 submat = submats[0]; 1720 if (submat->rmap->n != nrow || submat->cmap->n != ncol) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Cannot reuse matrix. wrong size"); 1721 1722 subc = (Mat_SeqAIJ*)submat->data; 1723 rmax = subc->rmax; 1724 smatis1 = subc->submatis1; 1725 nrqs = smatis1->nrqs; 1726 nrqr = smatis1->nrqr; 1727 rbuf1 = smatis1->rbuf1; 1728 rbuf2 = smatis1->rbuf2; 1729 rbuf3 = smatis1->rbuf3; 1730 req_source2 = smatis1->req_source2; 1731 1732 sbuf1 = smatis1->sbuf1; 1733 sbuf2 = smatis1->sbuf2; 1734 ptr = smatis1->ptr; 1735 tmp = smatis1->tmp; 1736 ctr = smatis1->ctr; 1737 1738 pa = smatis1->pa; 1739 req_size = smatis1->req_size; 1740 req_source1 = smatis1->req_source1; 1741 1742 allcolumns = smatis1->allcolumns; 1743 row2proc = smatis1->row2proc; 1744 rmap = smatis1->rmap; 1745 cmap = smatis1->cmap; 1746 #if defined(PETSC_USE_CTABLE) 1747 rmap_loc = smatis1->rmap_loc; 1748 cmap_loc = smatis1->cmap_loc; 1749 #endif 1750 } 1751 1752 /* Post recv matrix values */ 1753 ierr = PetscMalloc3(nrqs+1,&rbuf4, rmax,&subcols, rmax,&subvals);CHKERRQ(ierr); 1754 ierr = PetscMalloc4(nrqs+1,&r_waits4,nrqr+1,&s_waits4,nrqs+1,&r_status4,nrqr+1,&s_status4);CHKERRQ(ierr); 1755 ierr = PetscObjectGetNewTag((PetscObject)C,&tag4);CHKERRQ(ierr); 1756 for (i=0; i<nrqs; ++i) { 1757 ierr = PetscMalloc1(rbuf2[i][0]+1,&rbuf4[i]);CHKERRQ(ierr); 1758 ierr = MPI_Irecv(rbuf4[i],rbuf2[i][0],MPIU_SCALAR,req_source2[i],tag4,comm,r_waits4+i);CHKERRQ(ierr); 1759 } 1760 1761 /* Allocate sending buffers for a->a, and send them off */ 1762 ierr = PetscMalloc1(nrqr+1,&sbuf_aa);CHKERRQ(ierr); 1763 for (i=0,j=0; i<nrqr; i++) j += req_size[i]; 1764 ierr = PetscMalloc1(j+1,&sbuf_aa[0]);CHKERRQ(ierr); 1765 for (i=1; i<nrqr; i++) sbuf_aa[i] = sbuf_aa[i-1] + req_size[i-1]; 1766 1767 for (i=0; i<nrqr; i++) { 1768 rbuf1_i = rbuf1[i]; 1769 sbuf_aa_i = sbuf_aa[i]; 1770 ct1 = 2*rbuf1_i[0]+1; 1771 ct2 = 0; 1772 /* max1=rbuf1_i[0]; if (max1 != 1) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"max1 !=1"); */ 1773 1774 kmax = rbuf1_i[2]; 1775 for (k=0; k<kmax; k++,ct1++) { 1776 row = rbuf1_i[ct1] - rstart; 1777 nzA = ai[row+1] - ai[row]; 1778 nzB = bi[row+1] - bi[row]; 1779 ncols = nzA + nzB; 1780 cworkB = bj + bi[row]; 1781 vworkA = a_a + ai[row]; 1782 vworkB = b_a + bi[row]; 1783 1784 /* load the column values for this row into vals*/ 1785 vals = sbuf_aa_i + ct2; 1786 1787 lwrite = 0; 1788 for (l=0; l<nzB; l++) { 1789 if ((bmap[cworkB[l]]) < cstart) vals[lwrite++] = vworkB[l]; 1790 } 1791 for (l=0; l<nzA; l++) vals[lwrite++] = vworkA[l]; 1792 for (l=0; l<nzB; l++) { 1793 if ((bmap[cworkB[l]]) >= cend) vals[lwrite++] = vworkB[l]; 1794 } 1795 1796 ct2 += ncols; 1797 } 1798 ierr = MPI_Isend(sbuf_aa_i,req_size[i],MPIU_SCALAR,req_source1[i],tag4,comm,s_waits4+i);CHKERRQ(ierr); 1799 } 1800 1801 /* Assemble submat */ 1802 /* First assemble the local rows */ 1803 for (j=0; j<nrow; j++) { 1804 row = irow[j]; 1805 proc = row2proc[j]; 1806 if (proc == rank) { 1807 Crow = row - rstart; /* local row index of C */ 1808 #if defined(PETSC_USE_CTABLE) 1809 row = rmap_loc[Crow]; /* row index of submat */ 1810 #else 1811 row = rmap[row]; 1812 #endif 1813 1814 if (allcolumns) { 1815 /* diagonal part A = c->A */ 1816 ncols = ai[Crow+1] - ai[Crow]; 1817 cols = aj + ai[Crow]; 1818 vals = a->a + ai[Crow]; 1819 i = 0; 1820 for (k=0; k<ncols; k++) { 1821 subcols[i] = cols[k] + cstart; 1822 subvals[i++] = vals[k]; 1823 } 1824 1825 /* off-diagonal part B = c->B */ 1826 ncols = bi[Crow+1] - bi[Crow]; 1827 cols = bj + bi[Crow]; 1828 vals = b->a + bi[Crow]; 1829 for (k=0; k<ncols; k++) { 1830 subcols[i] = bmap[cols[k]]; 1831 subvals[i++] = vals[k]; 1832 } 1833 1834 ierr = MatSetValues_SeqAIJ(submat,1,&row,i,subcols,subvals,INSERT_VALUES);CHKERRQ(ierr); 1835 1836 } else { /* !allcolumns */ 1837 #if defined(PETSC_USE_CTABLE) 1838 /* diagonal part A = c->A */ 1839 ncols = ai[Crow+1] - ai[Crow]; 1840 cols = aj + ai[Crow]; 1841 vals = a->a + ai[Crow]; 1842 i = 0; 1843 for (k=0; k<ncols; k++) { 1844 tcol = cmap_loc[cols[k]]; 1845 if (tcol) { 1846 subcols[i] = --tcol; 1847 subvals[i++] = vals[k]; 1848 } 1849 } 1850 1851 /* off-diagonal part B = c->B */ 1852 ncols = bi[Crow+1] - bi[Crow]; 1853 cols = bj + bi[Crow]; 1854 vals = b->a + bi[Crow]; 1855 for (k=0; k<ncols; k++) { 1856 ierr = PetscTableFind(cmap,bmap[cols[k]]+1,&tcol);CHKERRQ(ierr); 1857 if (tcol) { 1858 subcols[i] = --tcol; 1859 subvals[i++] = vals[k]; 1860 } 1861 } 1862 #else 1863 /* diagonal part A = c->A */ 1864 ncols = ai[Crow+1] - ai[Crow]; 1865 cols = aj + ai[Crow]; 1866 vals = a->a + ai[Crow]; 1867 i = 0; 1868 for (k=0; k<ncols; k++) { 1869 tcol = cmap[cols[k]+cstart]; 1870 if (tcol) { 1871 subcols[i] = --tcol; 1872 subvals[i++] = vals[k]; 1873 } 1874 } 1875 1876 /* off-diagonal part B = c->B */ 1877 ncols = bi[Crow+1] - bi[Crow]; 1878 cols = bj + bi[Crow]; 1879 vals = b->a + bi[Crow]; 1880 for (k=0; k<ncols; k++) { 1881 tcol = cmap[bmap[cols[k]]]; 1882 if (tcol) { 1883 subcols[i] = --tcol; 1884 subvals[i++] = vals[k]; 1885 } 1886 } 1887 #endif 1888 ierr = MatSetValues_SeqAIJ(submat,1,&row,i,subcols,subvals,INSERT_VALUES);CHKERRQ(ierr); 1889 } 1890 } 1891 } 1892 1893 /* Now assemble the off-proc rows */ 1894 for (i=0; i<nrqs; i++) { /* for each requested message */ 1895 /* recv values from other processes */ 1896 ierr = MPI_Waitany(nrqs,r_waits4,&idex,r_status4+i);CHKERRQ(ierr); 1897 proc = pa[idex]; 1898 sbuf1_i = sbuf1[proc]; 1899 /* jmax = sbuf1_i[0]; if (jmax != 1)SETERRQ1(PETSC_COMM_SELF,0,"jmax %d != 1",jmax); */ 1900 ct1 = 2 + 1; 1901 ct2 = 0; /* count of received C->j */ 1902 ct3 = 0; /* count of received C->j that will be inserted into submat */ 1903 rbuf2_i = rbuf2[idex]; /* int** received length of C->j from other processes */ 1904 rbuf3_i = rbuf3[idex]; /* int** received C->j from other processes */ 1905 rbuf4_i = rbuf4[idex]; /* scalar** received C->a from other processes */ 1906 1907 /* is_no = sbuf1_i[2*j-1]; if (is_no != 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"is_no !=0"); */ 1908 max1 = sbuf1_i[2]; /* num of rows */ 1909 for (k=0; k<max1; k++,ct1++) { /* for each recved row */ 1910 row = sbuf1_i[ct1]; /* row index of submat */ 1911 if (!allcolumns) { 1912 idex = 0; 1913 if (scall == MAT_INITIAL_MATRIX) { 1914 nnz = rbuf2_i[ct1]; /* num of C entries in this row */ 1915 for (l=0; l<nnz; l++,ct2++) { /* for each recved column */ 1916 #if defined(PETSC_USE_CTABLE) 1917 if (rbuf3_i[ct2] >= cstart && rbuf3_i[ct2] <cend) { 1918 tcol = cmap_loc[rbuf3_i[ct2] - cstart]; 1919 } else { 1920 ierr = PetscTableFind(cmap,rbuf3_i[ct2]+1,&tcol);CHKERRQ(ierr); 1921 } 1922 #else 1923 tcol = cmap[rbuf3_i[ct2]]; 1924 #endif 1925 if (tcol) { 1926 subcols[idex] = --tcol; 1927 subvals[idex++] = rbuf4_i[ct2]; 1928 1929 /* We receive an entire column of C, but a subset of it needs to be inserted into submat. 1930 For reuse, we replace received C->j with index that should be inserted to submat */ 1931 rbuf3_i[ct3++] = ct2; 1932 } 1933 } 1934 ierr = MatSetValues_SeqAIJ(submat,1,&row,idex,subcols,subvals,INSERT_VALUES);CHKERRQ(ierr); 1935 1936 } else { /* scall == MAT_REUSE_MATRIX */ 1937 submat = submats[0]; 1938 subc = (Mat_SeqAIJ*)submat->data; 1939 1940 nnz = subc->i[row+1] - subc->i[row]; /* num of submat entries in this row */ 1941 for (l=0; l<nnz; l++) { 1942 ct2 = rbuf3_i[ct3++]; /* index of rbuf4_i[] which needs to be inserted into submat */ 1943 subvals[idex++] = rbuf4_i[ct2]; 1944 } 1945 1946 bj = subc->j + subc->i[row]; 1947 ierr = MatSetValues_SeqAIJ(submat,1,&row,nnz,bj,subvals,INSERT_VALUES);CHKERRQ(ierr); 1948 } 1949 } else { /* allcolumns */ 1950 nnz = rbuf2_i[ct1]; /* num of C entries in this row */ 1951 ierr = MatSetValues_SeqAIJ(submat,1,&row,nnz,rbuf3_i+ct2,rbuf4_i+ct2,INSERT_VALUES);CHKERRQ(ierr); 1952 ct2 += nnz; 1953 } 1954 } 1955 } 1956 1957 /* sending a->a are done */ 1958 ierr = MPI_Waitall(nrqr,s_waits4,s_status4);CHKERRQ(ierr); 1959 ierr = PetscFree4(r_waits4,s_waits4,r_status4,s_status4);CHKERRQ(ierr); 1960 1961 ierr = MatAssemblyBegin(submat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1962 ierr = MatAssemblyEnd(submat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1963 submats[0] = submat; 1964 1965 /* Restore the indices */ 1966 ierr = ISRestoreIndices(isrow[0],&irow);CHKERRQ(ierr); 1967 if (!allcolumns) { 1968 ierr = ISRestoreIndices(iscol[0],&icol);CHKERRQ(ierr); 1969 } 1970 1971 /* Destroy allocated memory */ 1972 for (i=0; i<nrqs; ++i) { 1973 ierr = PetscFree3(rbuf4[i],subcols,subvals);CHKERRQ(ierr); 1974 } 1975 ierr = PetscFree3(rbuf4,subcols,subvals);CHKERRQ(ierr); 1976 ierr = PetscFree(sbuf_aa[0]);CHKERRQ(ierr); 1977 ierr = PetscFree(sbuf_aa);CHKERRQ(ierr); 1978 1979 if (scall == MAT_INITIAL_MATRIX) { 1980 ierr = PetscFree(lens);CHKERRQ(ierr); 1981 ierr = PetscFree(sbuf_aj[0]);CHKERRQ(ierr); 1982 ierr = PetscFree(sbuf_aj);CHKERRQ(ierr); 1983 } 1984 PetscFunctionReturn(0); 1985 } 1986 1987 PetscErrorCode MatGetSubMatrices_MPIAIJ_SingleIS(Mat C,PetscInt ismax,const IS isrow[],const IS iscol[],MatReuse scall,Mat *submat[]) 1988 { 1989 PetscErrorCode ierr; 1990 PetscInt ncol; 1991 PetscBool colflag,allcolumns=PETSC_FALSE; 1992 1993 PetscFunctionBegin; 1994 /* Allocate memory to hold all the submatrices */ 1995 if (scall == MAT_INITIAL_MATRIX) { 1996 ierr = PetscMalloc1(1,submat);CHKERRQ(ierr); 1997 } 1998 1999 /* Check for special case: each processor gets entire matrix columns */ 2000 ierr = ISIdentity(iscol[0],&colflag);CHKERRQ(ierr); 2001 ierr = ISGetLocalSize(iscol[0],&ncol);CHKERRQ(ierr); 2002 if (colflag && ncol == C->cmap->N) allcolumns = PETSC_TRUE; 2003 2004 ierr = MatGetSubMatrices_MPIAIJ_SingleIS_Local(C,ismax,isrow,iscol,scall,allcolumns,*submat);CHKERRQ(ierr); 2005 PetscFunctionReturn(0); 2006 } 2007 2008 PetscErrorCode MatGetSubMatrices_MPIAIJ(Mat C,PetscInt ismax,const IS isrow[],const IS iscol[],MatReuse scall,Mat *submat[]) 2009 { 2010 PetscErrorCode ierr; 2011 PetscInt nmax,nstages,i,pos,max_no,nrow,ncol,in[2],out[2]; 2012 PetscBool rowflag,colflag,wantallmatrix=PETSC_FALSE,*allcolumns; 2013 2014 PetscFunctionBegin; 2015 #if 0 2016 /* Check for special case: each processor gets entire matrix */ 2017 if (C->submat_singleis) { /* flag is set in PCSetUp_ASM() to skip several MPIU_Allreduce() */ 2018 ierr = MatGetSubMatrices_MPIAIJ_SingleIS(C,ismax,isrow,iscol,scall,submat);CHKERRQ(ierr); 2019 PetscFunctionReturn(0); 2020 } 2021 #endif 2022 2023 if (ismax == 1 && C->rmap->N == C->cmap->N) { 2024 ierr = ISIdentity(*isrow,&rowflag);CHKERRQ(ierr); 2025 ierr = ISIdentity(*iscol,&colflag);CHKERRQ(ierr); 2026 ierr = ISGetLocalSize(*isrow,&nrow);CHKERRQ(ierr); 2027 ierr = ISGetLocalSize(*iscol,&ncol);CHKERRQ(ierr); 2028 if (rowflag && colflag && nrow == C->rmap->N && ncol == C->cmap->N) { 2029 wantallmatrix = PETSC_TRUE; 2030 2031 ierr = PetscOptionsGetBool(((PetscObject)C)->options,((PetscObject)C)->prefix,"-use_fast_submatrix",&wantallmatrix,NULL);CHKERRQ(ierr); 2032 } 2033 } 2034 2035 /* Determine the number of stages through which submatrices are done 2036 Each stage will extract nmax submatrices. 2037 nmax is determined by the matrix column dimension. 2038 If the original matrix has 20M columns, only one submatrix per stage is allowed, etc. 2039 */ 2040 nmax = 20*1000000 / (C->cmap->N * sizeof(PetscInt)); 2041 if (!nmax) nmax = 1; 2042 nstages = ismax/nmax + ((ismax % nmax) ? 1 : 0); /* local nstages */ 2043 2044 /* Collect global wantallmatrix and nstages */ 2045 in[0] = -1*wantallmatrix; 2046 in[1] = nstages; 2047 ierr = MPIU_Allreduce(in,out,2,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)C));CHKERRQ(ierr); 2048 wantallmatrix = (PetscBool)(-out[0]); 2049 nstages = out[1]; /* Make sure every processor loops through the global nstages */ 2050 2051 if (wantallmatrix) { 2052 ierr = MatGetSubMatrix_MPIAIJ_All(C,MAT_GET_VALUES,scall,submat);CHKERRQ(ierr); 2053 PetscFunctionReturn(0); 2054 } 2055 2056 /* Allocate memory to hold all the submatrices */ 2057 if (scall == MAT_INITIAL_MATRIX) { 2058 ierr = PetscMalloc1(ismax+1,submat);CHKERRQ(ierr); 2059 } 2060 2061 if (scall == MAT_REUSE_MATRIX && !ismax) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"n=0 is not supported for MatGetSubMatrices(mat,n,isrow,iscol,MAT_REUSE_MATRIX,...). Set n=1 with zero-length isrow and iscolumn instead"); 2062 2063 /* Check for special case: each processor gets entire matrix columns */ 2064 ierr = PetscMalloc1(ismax+1,&allcolumns);CHKERRQ(ierr); 2065 for (i=0; i<ismax; i++) { 2066 ierr = ISIdentity(iscol[i],&colflag);CHKERRQ(ierr); 2067 ierr = ISGetLocalSize(iscol[i],&ncol);CHKERRQ(ierr); 2068 if (colflag && ncol == C->cmap->N) { 2069 allcolumns[i] = PETSC_TRUE; 2070 } else { 2071 allcolumns[i] = PETSC_FALSE; 2072 } 2073 } 2074 2075 for (i=0,pos=0; i<nstages; i++) { 2076 if (pos+nmax <= ismax) max_no = nmax; 2077 else if (pos == ismax) max_no = 0; 2078 else max_no = ismax-pos; 2079 ierr = MatGetSubMatrices_MPIAIJ_Local(C,max_no,isrow+pos,iscol+pos,scall,allcolumns+pos,*submat+pos);CHKERRQ(ierr); 2080 pos += max_no; 2081 } 2082 2083 ierr = PetscFree(allcolumns);CHKERRQ(ierr); 2084 PetscFunctionReturn(0); 2085 } 2086 2087 /* -------------------------------------------------------------------------*/ 2088 PetscErrorCode MatGetSubMatrices_MPIAIJ_Local(Mat C,PetscInt ismax,const IS isrow[],const IS iscol[],MatReuse scall,PetscBool *allcolumns,Mat *submats) 2089 { 2090 Mat_MPIAIJ *c = (Mat_MPIAIJ*)C->data; 2091 Mat A = c->A; 2092 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data,*b = (Mat_SeqAIJ*)c->B->data,*subc; 2093 const PetscInt **icol,**irow; 2094 PetscInt *nrow,*ncol,start; 2095 PetscErrorCode ierr; 2096 PetscMPIInt rank,size,tag0,tag2,tag3,tag4,*w1,*w2,*w3,*w4,nrqr; 2097 PetscInt **sbuf1,**sbuf2,i,j,k,l,ct1,ct2,**rbuf1,row,proc=-1; 2098 PetscInt nrqs=0,msz,**ptr=NULL,*req_size=NULL,*ctr=NULL,*pa,*tmp=NULL,tcol; 2099 PetscInt **rbuf3=NULL,*req_source1=NULL,*req_source2,**sbuf_aj,**rbuf2=NULL,max1,max2; 2100 PetscInt **lens,is_no,ncols,*cols,mat_i,*mat_j,tmp2,jmax; 2101 #if defined(PETSC_USE_CTABLE) 2102 PetscTable *cmap,cmap_i=NULL,*rmap,rmap_i; 2103 #else 2104 PetscInt **cmap,*cmap_i=NULL,**rmap,*rmap_i; 2105 #endif 2106 const PetscInt *irow_i; 2107 PetscInt ctr_j,*sbuf1_j,*sbuf_aj_i,*rbuf1_i,kmax,*lens_i; 2108 MPI_Request *s_waits1,*r_waits1,*s_waits2,*r_waits2,*r_waits3; 2109 MPI_Request *r_waits4,*s_waits3,*s_waits4; 2110 MPI_Status *r_status1,*r_status2,*s_status1,*s_status3,*s_status2; 2111 MPI_Status *r_status3,*r_status4,*s_status4; 2112 MPI_Comm comm; 2113 PetscScalar **rbuf4,**sbuf_aa,*vals,*mat_a,*sbuf_aa_i; 2114 PetscMPIInt *onodes1,*olengths1,end; 2115 PetscInt **row2proc,*row2proc_i; 2116 Mat_SubMat **smats,*smat_i; 2117 PetscBool *issorted; 2118 2119 PetscFunctionBegin; 2120 ierr = PetscObjectGetComm((PetscObject)C,&comm);CHKERRQ(ierr); 2121 size = c->size; 2122 rank = c->rank; 2123 2124 ierr = PetscMalloc5(ismax,&irow,ismax,&icol,ismax,&nrow,ismax,&ncol,ismax,&issorted);CHKERRQ(ierr); 2125 for (i=0; i<ismax; i++) { 2126 ierr = ISSorted(isrow[i],&issorted[i]);CHKERRQ(ierr); 2127 2128 ierr = ISGetIndices(isrow[i],&irow[i]);CHKERRQ(ierr); 2129 ierr = ISGetLocalSize(isrow[i],&nrow[i]);CHKERRQ(ierr); 2130 if (allcolumns[i]) { 2131 icol[i] = NULL; 2132 ncol[i] = C->cmap->N; 2133 } else { 2134 ierr = ISGetIndices(iscol[i],&icol[i]);CHKERRQ(ierr); 2135 ierr = ISGetLocalSize(iscol[i],&ncol[i]);CHKERRQ(ierr); 2136 } 2137 } 2138 2139 ierr = PetscMalloc1(ismax,&smats);CHKERRQ(ierr); 2140 ierr = PetscMalloc1(ismax,&row2proc);CHKERRQ(ierr); 2141 ierr = PetscMalloc2(ismax,&cmap,ismax,&rmap);CHKERRQ(ierr); 2142 2143 if (scall == MAT_REUSE_MATRIX) { 2144 /* Assumes new rows are same length as the old rows */ 2145 for (i=0; i<ismax; i++) { 2146 subc = (Mat_SeqAIJ*)(submats[i]->data); 2147 if ((submats[i]->rmap->n != nrow[i]) || (submats[i]->cmap->n != ncol[i])) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Cannot reuse matrix. wrong size"); 2148 2149 /* Initial matrix as if empty */ 2150 ierr = PetscMemzero(subc->ilen,submats[i]->rmap->n*sizeof(PetscInt));CHKERRQ(ierr); 2151 2152 /* Initial matrix as if empty */ 2153 submats[i]->factortype = C->factortype; 2154 2155 smat_i = subc->submatis1; 2156 smats[i] = smat_i; 2157 2158 nrqs = smat_i->nrqs; 2159 nrqr = smat_i->nrqr; 2160 rbuf1 = smat_i->rbuf1; 2161 rbuf2 = smat_i->rbuf2; 2162 rbuf3 = smat_i->rbuf3; 2163 req_source2 = smat_i->req_source2; 2164 2165 sbuf1 = smat_i->sbuf1; 2166 sbuf2 = smat_i->sbuf2; 2167 ptr = smat_i->ptr; 2168 tmp = smat_i->tmp; 2169 ctr = smat_i->ctr; 2170 2171 pa = smat_i->pa; 2172 req_size = smat_i->req_size; 2173 req_source1 = smat_i->req_source1; 2174 2175 allcolumns[i] = smat_i->allcolumns; 2176 row2proc[i] = smat_i->row2proc; 2177 rmap[i] = smat_i->rmap; 2178 cmap[i] = smat_i->cmap; 2179 } 2180 } else { /* scall == MAT_INITIAL_MATRIX */ 2181 /* Get some new tags to keep the communication clean */ 2182 ierr = PetscObjectGetNewTag((PetscObject)C,&tag2);CHKERRQ(ierr); 2183 ierr = PetscObjectGetNewTag((PetscObject)C,&tag3);CHKERRQ(ierr); 2184 2185 /* evaluate communication - mesg to who, length of mesg, and buffer space 2186 required. Based on this, buffers are allocated, and data copied into them*/ 2187 ierr = PetscCalloc4(size,&w1,size,&w2,size,&w3,size,&w4);CHKERRQ(ierr); /* mesg size, initialize work vectors */ 2188 2189 for (i=0; i<ismax; i++) { 2190 jmax = nrow[i]; 2191 irow_i = irow[i]; 2192 2193 ierr = PetscMalloc1(jmax,&row2proc_i);CHKERRQ(ierr); 2194 row2proc[i] = row2proc_i; 2195 2196 if (issorted[i]) proc = 0; 2197 for (j=0; j<jmax; j++) { 2198 if (!issorted[i]) proc = 0; 2199 row = irow_i[j]; 2200 while (row >= C->rmap->range[proc+1]) proc++; 2201 w4[proc]++; 2202 row2proc_i[j] = proc; /* map row index to proc */ 2203 } 2204 for (j=0; j<size; j++) { 2205 if (w4[j]) { w1[j] += w4[j]; w3[j]++; w4[j] = 0;} 2206 } 2207 } 2208 2209 nrqs = 0; /* no of outgoing messages */ 2210 msz = 0; /* total mesg length (for all procs) */ 2211 w1[rank] = 0; /* no mesg sent to self */ 2212 w3[rank] = 0; 2213 for (i=0; i<size; i++) { 2214 if (w1[i]) { w2[i] = 1; nrqs++;} /* there exists a message to proc i */ 2215 } 2216 ierr = PetscMalloc1(nrqs+1,&pa);CHKERRQ(ierr); /*(proc -array)*/ 2217 for (i=0,j=0; i<size; i++) { 2218 if (w1[i]) { pa[j] = i; j++; } 2219 } 2220 2221 /* Each message would have a header = 1 + 2*(no of IS) + data */ 2222 for (i=0; i<nrqs; i++) { 2223 j = pa[i]; 2224 w1[j] += w2[j] + 2* w3[j]; 2225 msz += w1[j]; 2226 } 2227 ierr = PetscInfo2(0,"Number of outgoing messages %D Total message length %D\n",nrqs,msz);CHKERRQ(ierr); 2228 2229 /* Determine the number of messages to expect, their lengths, from from-ids */ 2230 ierr = PetscGatherNumberOfMessages(comm,w2,w1,&nrqr);CHKERRQ(ierr); 2231 ierr = PetscGatherMessageLengths(comm,nrqs,nrqr,w1,&onodes1,&olengths1);CHKERRQ(ierr); 2232 2233 /* Now post the Irecvs corresponding to these messages */ 2234 tag0 = ((PetscObject)C)->tag; 2235 ierr = PetscPostIrecvInt(comm,tag0,nrqr,onodes1,olengths1,&rbuf1,&r_waits1);CHKERRQ(ierr); 2236 2237 ierr = PetscFree(onodes1);CHKERRQ(ierr); 2238 ierr = PetscFree(olengths1);CHKERRQ(ierr); 2239 2240 /* Allocate Memory for outgoing messages */ 2241 ierr = PetscMalloc4(size,&sbuf1,size,&ptr,2*msz,&tmp,size,&ctr);CHKERRQ(ierr); 2242 ierr = PetscMemzero(sbuf1,size*sizeof(PetscInt*));CHKERRQ(ierr); 2243 ierr = PetscMemzero(ptr,size*sizeof(PetscInt*));CHKERRQ(ierr); 2244 2245 { 2246 PetscInt *iptr = tmp; 2247 k = 0; 2248 for (i=0; i<nrqs; i++) { 2249 j = pa[i]; 2250 iptr += k; 2251 sbuf1[j] = iptr; 2252 k = w1[j]; 2253 } 2254 } 2255 2256 /* Form the outgoing messages. Initialize the header space */ 2257 for (i=0; i<nrqs; i++) { 2258 j = pa[i]; 2259 sbuf1[j][0] = 0; 2260 ierr = PetscMemzero(sbuf1[j]+1,2*w3[j]*sizeof(PetscInt));CHKERRQ(ierr); 2261 ptr[j] = sbuf1[j] + 2*w3[j] + 1; 2262 } 2263 2264 /* Parse the isrow and copy data into outbuf */ 2265 for (i=0; i<ismax; i++) { 2266 row2proc_i = row2proc[i]; 2267 ierr = PetscMemzero(ctr,size*sizeof(PetscInt));CHKERRQ(ierr); 2268 irow_i = irow[i]; 2269 jmax = nrow[i]; 2270 for (j=0; j<jmax; j++) { /* parse the indices of each IS */ 2271 proc = row2proc_i[j]; 2272 if (proc != rank) { /* copy to the outgoing buf*/ 2273 ctr[proc]++; 2274 *ptr[proc] = irow_i[j]; 2275 ptr[proc]++; 2276 } 2277 } 2278 /* Update the headers for the current IS */ 2279 for (j=0; j<size; j++) { /* Can Optimise this loop too */ 2280 if ((ctr_j = ctr[j])) { 2281 sbuf1_j = sbuf1[j]; 2282 k = ++sbuf1_j[0]; 2283 sbuf1_j[2*k] = ctr_j; 2284 sbuf1_j[2*k-1] = i; 2285 } 2286 } 2287 } 2288 2289 /* Now post the sends */ 2290 ierr = PetscMalloc1(nrqs+1,&s_waits1);CHKERRQ(ierr); 2291 for (i=0; i<nrqs; ++i) { 2292 j = pa[i]; 2293 ierr = MPI_Isend(sbuf1[j],w1[j],MPIU_INT,j,tag0,comm,s_waits1+i);CHKERRQ(ierr); 2294 } 2295 2296 /* Post Receives to capture the buffer size */ 2297 ierr = PetscMalloc1(nrqs+1,&r_waits2);CHKERRQ(ierr); 2298 ierr = PetscMalloc3(nrqs+1,&req_source2,nrqs+1,&rbuf2,nrqs+1,&rbuf3);CHKERRQ(ierr); 2299 rbuf2[0] = tmp + msz; 2300 for (i=1; i<nrqs; ++i) { 2301 rbuf2[i] = rbuf2[i-1]+w1[pa[i-1]]; 2302 } 2303 for (i=0; i<nrqs; ++i) { 2304 j = pa[i]; 2305 ierr = MPI_Irecv(rbuf2[i],w1[j],MPIU_INT,j,tag2,comm,r_waits2+i);CHKERRQ(ierr); 2306 } 2307 2308 /* Send to other procs the buf size they should allocate */ 2309 /* Receive messages*/ 2310 ierr = PetscMalloc1(nrqr+1,&s_waits2);CHKERRQ(ierr); 2311 ierr = PetscMalloc1(nrqr+1,&r_status1);CHKERRQ(ierr); 2312 ierr = PetscMalloc3(nrqr,&sbuf2,nrqr,&req_size,nrqr,&req_source1);CHKERRQ(ierr); 2313 { 2314 PetscInt *sAi = a->i,*sBi = b->i,id,rstart = C->rmap->rstart; 2315 PetscInt *sbuf2_i; 2316 2317 ierr = MPI_Waitall(nrqr,r_waits1,r_status1);CHKERRQ(ierr); 2318 for (i=0; i<nrqr; ++i) { 2319 req_size[i] = 0; 2320 rbuf1_i = rbuf1[i]; 2321 start = 2*rbuf1_i[0] + 1; 2322 ierr = MPI_Get_count(r_status1+i,MPIU_INT,&end);CHKERRQ(ierr); 2323 ierr = PetscMalloc1(end+1,&sbuf2[i]);CHKERRQ(ierr); 2324 sbuf2_i = sbuf2[i]; 2325 for (j=start; j<end; j++) { 2326 id = rbuf1_i[j] - rstart; 2327 ncols = sAi[id+1] - sAi[id] + sBi[id+1] - sBi[id]; 2328 sbuf2_i[j] = ncols; 2329 req_size[i] += ncols; 2330 } 2331 req_source1[i] = r_status1[i].MPI_SOURCE; 2332 /* form the header */ 2333 sbuf2_i[0] = req_size[i]; 2334 for (j=1; j<start; j++) sbuf2_i[j] = rbuf1_i[j]; 2335 2336 ierr = MPI_Isend(sbuf2_i,end,MPIU_INT,req_source1[i],tag2,comm,s_waits2+i);CHKERRQ(ierr); 2337 } 2338 } 2339 ierr = PetscFree(r_status1);CHKERRQ(ierr); 2340 ierr = PetscFree(r_waits1);CHKERRQ(ierr); 2341 ierr = PetscFree4(w1,w2,w3,w4);CHKERRQ(ierr); 2342 2343 /* Receive messages*/ 2344 ierr = PetscMalloc1(nrqs+1,&r_waits3);CHKERRQ(ierr); 2345 ierr = PetscMalloc1(nrqs+1,&r_status2);CHKERRQ(ierr); 2346 2347 ierr = MPI_Waitall(nrqs,r_waits2,r_status2);CHKERRQ(ierr); 2348 for (i=0; i<nrqs; ++i) { 2349 ierr = PetscMalloc1(rbuf2[i][0]+1,&rbuf3[i]);CHKERRQ(ierr); 2350 req_source2[i] = r_status2[i].MPI_SOURCE; 2351 ierr = MPI_Irecv(rbuf3[i],rbuf2[i][0],MPIU_INT,req_source2[i],tag3,comm,r_waits3+i);CHKERRQ(ierr); 2352 } 2353 ierr = PetscFree(r_status2);CHKERRQ(ierr); 2354 ierr = PetscFree(r_waits2);CHKERRQ(ierr); 2355 2356 /* Wait on sends1 and sends2 */ 2357 ierr = PetscMalloc1(nrqs+1,&s_status1);CHKERRQ(ierr); 2358 ierr = PetscMalloc1(nrqr+1,&s_status2);CHKERRQ(ierr); 2359 2360 if (nrqs) {ierr = MPI_Waitall(nrqs,s_waits1,s_status1);CHKERRQ(ierr);} 2361 if (nrqr) {ierr = MPI_Waitall(nrqr,s_waits2,s_status2);CHKERRQ(ierr);} 2362 ierr = PetscFree(s_status1);CHKERRQ(ierr); 2363 ierr = PetscFree(s_status2);CHKERRQ(ierr); 2364 ierr = PetscFree(s_waits1);CHKERRQ(ierr); 2365 ierr = PetscFree(s_waits2);CHKERRQ(ierr); 2366 2367 /* Now allocate sending buffers for a->j, and send them off */ 2368 ierr = PetscMalloc1(nrqr+1,&sbuf_aj);CHKERRQ(ierr); 2369 for (i=0,j=0; i<nrqr; i++) j += req_size[i]; 2370 ierr = PetscMalloc1(j+1,&sbuf_aj[0]);CHKERRQ(ierr); 2371 for (i=1; i<nrqr; i++) sbuf_aj[i] = sbuf_aj[i-1] + req_size[i-1]; 2372 2373 ierr = PetscMalloc1(nrqr+1,&s_waits3);CHKERRQ(ierr); 2374 { 2375 PetscInt nzA,nzB,*a_i = a->i,*b_i = b->i,lwrite; 2376 PetscInt *cworkA,*cworkB,cstart = C->cmap->rstart,rstart = C->rmap->rstart,*bmap = c->garray; 2377 PetscInt cend = C->cmap->rend; 2378 PetscInt *a_j = a->j,*b_j = b->j,ctmp; 2379 2380 for (i=0; i<nrqr; i++) { 2381 rbuf1_i = rbuf1[i]; 2382 sbuf_aj_i = sbuf_aj[i]; 2383 ct1 = 2*rbuf1_i[0] + 1; 2384 ct2 = 0; 2385 for (j=1,max1=rbuf1_i[0]; j<=max1; j++) { 2386 kmax = rbuf1[i][2*j]; 2387 for (k=0; k<kmax; k++,ct1++) { 2388 row = rbuf1_i[ct1] - rstart; 2389 nzA = a_i[row+1] - a_i[row]; nzB = b_i[row+1] - b_i[row]; 2390 ncols = nzA + nzB; 2391 cworkA = a_j + a_i[row]; cworkB = b_j + b_i[row]; 2392 2393 /* load the column indices for this row into sorted cols */ 2394 cols = sbuf_aj_i + ct2; 2395 2396 lwrite = 0; 2397 for (l=0; l<nzB; l++) { 2398 if ((ctmp = bmap[cworkB[l]]) < cstart) cols[lwrite++] = ctmp; 2399 } 2400 for (l=0; l<nzA; l++) cols[lwrite++] = cstart + cworkA[l]; 2401 for (l=0; l<nzB; l++) { 2402 if ((ctmp = bmap[cworkB[l]]) >= cend) cols[lwrite++] = ctmp; 2403 } 2404 2405 ct2 += ncols; 2406 } 2407 } 2408 ierr = MPI_Isend(sbuf_aj_i,req_size[i],MPIU_INT,req_source1[i],tag3,comm,s_waits3+i);CHKERRQ(ierr); 2409 } 2410 } 2411 ierr = PetscMalloc1(nrqs+1,&r_status3);CHKERRQ(ierr); 2412 ierr = PetscMalloc1(nrqr+1,&s_status3);CHKERRQ(ierr); 2413 2414 /* create col map: global col of C -> local col of submatrices */ 2415 { 2416 const PetscInt *icol_i; 2417 #if defined(PETSC_USE_CTABLE) 2418 for (i=0; i<ismax; i++) { 2419 if (!allcolumns[i]) { 2420 ierr = PetscTableCreate(ncol[i]+1,C->cmap->N+1,&cmap[i]);CHKERRQ(ierr); 2421 2422 jmax = ncol[i]; 2423 icol_i = icol[i]; 2424 cmap_i = cmap[i]; 2425 for (j=0; j<jmax; j++) { 2426 ierr = PetscTableAdd(cmap[i],icol_i[j]+1,j+1,INSERT_VALUES);CHKERRQ(ierr); 2427 } 2428 } else cmap[i] = NULL; 2429 } 2430 #else 2431 for (i=0; i<ismax; i++) { 2432 if (!allcolumns[i]) { 2433 ierr = PetscCalloc1(C->cmap->N,&cmap[i]);CHKERRQ(ierr); 2434 jmax = ncol[i]; 2435 icol_i = icol[i]; 2436 cmap_i = cmap[i]; 2437 for (j=0; j<jmax; j++) { 2438 cmap_i[icol_i[j]] = j+1; 2439 } 2440 } else cmap[i] = NULL; 2441 } 2442 #endif 2443 } 2444 2445 /* Create lens which is required for MatCreate... */ 2446 for (i=0,j=0; i<ismax; i++) j += nrow[i]; 2447 ierr = PetscMalloc1(ismax,&lens);CHKERRQ(ierr); 2448 2449 if (ismax) { 2450 ierr = PetscCalloc1(j,&lens[0]);CHKERRQ(ierr); 2451 } 2452 for (i=1; i<ismax; i++) lens[i] = lens[i-1] + nrow[i-1]; 2453 2454 /* Update lens from local data */ 2455 for (i=0; i<ismax; i++) { 2456 row2proc_i = row2proc[i]; 2457 jmax = nrow[i]; 2458 if (!allcolumns[i]) cmap_i = cmap[i]; 2459 irow_i = irow[i]; 2460 lens_i = lens[i]; 2461 for (j=0; j<jmax; j++) { 2462 row = irow_i[j]; 2463 proc = row2proc_i[j]; 2464 if (proc == rank) { 2465 ierr = MatGetRow_MPIAIJ(C,row,&ncols,&cols,0);CHKERRQ(ierr); 2466 if (!allcolumns[i]) { 2467 for (k=0; k<ncols; k++) { 2468 #if defined(PETSC_USE_CTABLE) 2469 ierr = PetscTableFind(cmap_i,cols[k]+1,&tcol);CHKERRQ(ierr); 2470 #else 2471 tcol = cmap_i[cols[k]]; 2472 #endif 2473 if (tcol) lens_i[j]++; 2474 } 2475 } else { /* allcolumns */ 2476 lens_i[j] = ncols; 2477 } 2478 ierr = MatRestoreRow_MPIAIJ(C,row,&ncols,&cols,0);CHKERRQ(ierr); 2479 } 2480 } 2481 } 2482 2483 /* Create row map: global row of C -> local row of submatrices */ 2484 #if defined(PETSC_USE_CTABLE) 2485 for (i=0; i<ismax; i++) { 2486 ierr = PetscTableCreate(nrow[i]+1,C->rmap->N+1,&rmap[i]);CHKERRQ(ierr); 2487 irow_i = irow[i]; 2488 jmax = nrow[i]; 2489 for (j=0; j<jmax; j++) { 2490 ierr = PetscTableAdd(rmap[i],irow_i[j]+1,j+1,INSERT_VALUES);CHKERRQ(ierr); 2491 } 2492 } 2493 #else 2494 for (i=0; i<ismax; i++) { 2495 ierr = PetscCalloc1(C->rmap->N,&rmap[i]);CHKERRQ(ierr); 2496 rmap_i = rmap[i]; 2497 irow_i = irow[i]; 2498 jmax = nrow[i]; 2499 for (j=0; j<jmax; j++) { 2500 rmap_i[irow_i[j]] = j; 2501 } 2502 } 2503 #endif 2504 2505 /* Update lens from offproc data */ 2506 { 2507 PetscInt *rbuf2_i,*rbuf3_i,*sbuf1_i; 2508 2509 ierr = MPI_Waitall(nrqs,r_waits3,r_status3);CHKERRQ(ierr); 2510 for (tmp2=0; tmp2<nrqs; tmp2++) { 2511 sbuf1_i = sbuf1[pa[tmp2]]; 2512 jmax = sbuf1_i[0]; 2513 ct1 = 2*jmax+1; 2514 ct2 = 0; 2515 rbuf2_i = rbuf2[tmp2]; 2516 rbuf3_i = rbuf3[tmp2]; 2517 for (j=1; j<=jmax; j++) { 2518 is_no = sbuf1_i[2*j-1]; 2519 max1 = sbuf1_i[2*j]; 2520 lens_i = lens[is_no]; 2521 if (!allcolumns[is_no]) cmap_i = cmap[is_no]; 2522 rmap_i = rmap[is_no]; 2523 for (k=0; k<max1; k++,ct1++) { 2524 #if defined(PETSC_USE_CTABLE) 2525 ierr = PetscTableFind(rmap_i,sbuf1_i[ct1]+1,&row);CHKERRQ(ierr); 2526 row--; 2527 if (row < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"row not found in table"); 2528 #else 2529 row = rmap_i[sbuf1_i[ct1]]; /* the val in the new matrix to be */ 2530 #endif 2531 max2 = rbuf2_i[ct1]; 2532 for (l=0; l<max2; l++,ct2++) { 2533 if (!allcolumns[is_no]) { 2534 #if defined(PETSC_USE_CTABLE) 2535 ierr = PetscTableFind(cmap_i,rbuf3_i[ct2]+1,&tcol);CHKERRQ(ierr); 2536 #else 2537 tcol = cmap_i[rbuf3_i[ct2]]; 2538 #endif 2539 if (tcol) lens_i[row]++; 2540 } else { /* allcolumns */ 2541 lens_i[row]++; /* lens_i[row] += max2 ? */ 2542 } 2543 } 2544 } 2545 } 2546 } 2547 } 2548 ierr = PetscFree(r_status3);CHKERRQ(ierr); 2549 ierr = PetscFree(r_waits3);CHKERRQ(ierr); 2550 if (nrqr) {ierr = MPI_Waitall(nrqr,s_waits3,s_status3);CHKERRQ(ierr);} 2551 ierr = PetscFree(s_status3);CHKERRQ(ierr); 2552 ierr = PetscFree(s_waits3);CHKERRQ(ierr); 2553 2554 /* Create the submatrices */ 2555 for (i=0; i<ismax; i++) { 2556 PetscInt rbs,cbs; 2557 2558 ierr = ISGetBlockSize(isrow[i],&rbs);CHKERRQ(ierr); 2559 ierr = ISGetBlockSize(iscol[i],&cbs);CHKERRQ(ierr); 2560 2561 ierr = MatCreate(PETSC_COMM_SELF,submats+i);CHKERRQ(ierr); 2562 ierr = MatSetSizes(submats[i],nrow[i],ncol[i],PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 2563 2564 ierr = MatSetBlockSizes(submats[i],rbs,cbs);CHKERRQ(ierr); 2565 ierr = MatSetType(submats[i],((PetscObject)A)->type_name);CHKERRQ(ierr); 2566 ierr = MatSeqAIJSetPreallocation(submats[i],0,lens[i]);CHKERRQ(ierr); 2567 2568 /* create struct Mat_SubMat and attached it to submat */ 2569 ierr = PetscNew(&smat_i);CHKERRQ(ierr); 2570 subc = (Mat_SeqAIJ*)submats[i]->data; 2571 subc->submatis1 = smat_i; 2572 smats[i] = smat_i; 2573 2574 smat_i->destroy = submats[i]->ops->destroy; 2575 submats[i]->ops->destroy = MatDestroy_MPIAIJ_MatGetSubmatrices; 2576 submats[i]->factortype = C->factortype; 2577 2578 smat_i->id = i; 2579 smat_i->nrqs = nrqs; 2580 smat_i->nrqr = nrqr; 2581 smat_i->rbuf1 = rbuf1; 2582 smat_i->rbuf2 = rbuf2; 2583 smat_i->rbuf3 = rbuf3; 2584 smat_i->sbuf2 = sbuf2; 2585 smat_i->req_source2 = req_source2; 2586 2587 smat_i->sbuf1 = sbuf1; 2588 smat_i->ptr = ptr; 2589 smat_i->tmp = tmp; 2590 smat_i->ctr = ctr; 2591 2592 smat_i->pa = pa; 2593 smat_i->req_size = req_size; 2594 smat_i->req_source1 = req_source1; 2595 2596 smat_i->allcolumns = allcolumns[i]; 2597 smat_i->row2proc = row2proc[i]; 2598 smat_i->rmap = rmap[i]; 2599 smat_i->cmap = cmap[i]; 2600 } 2601 2602 if (ismax) {ierr = PetscFree(lens[0]);CHKERRQ(ierr);} 2603 ierr = PetscFree(lens);CHKERRQ(ierr); 2604 ierr = PetscFree(sbuf_aj[0]);CHKERRQ(ierr); 2605 ierr = PetscFree(sbuf_aj);CHKERRQ(ierr); 2606 2607 } /* endof scall == MAT_INITIAL_MATRIX */ 2608 2609 /* Post recv matrix values */ 2610 ierr = PetscObjectGetNewTag((PetscObject)C,&tag4);CHKERRQ(ierr); 2611 ierr = PetscMalloc1(nrqs+1,&rbuf4);CHKERRQ(ierr); 2612 ierr = PetscMalloc1(nrqs+1,&r_waits4);CHKERRQ(ierr); 2613 ierr = PetscMalloc1(nrqs+1,&r_status4);CHKERRQ(ierr); 2614 ierr = PetscMalloc1(nrqr+1,&s_status4);CHKERRQ(ierr); 2615 for (i=0; i<nrqs; ++i) { 2616 ierr = PetscMalloc1(rbuf2[i][0]+1,&rbuf4[i]);CHKERRQ(ierr); 2617 ierr = MPI_Irecv(rbuf4[i],rbuf2[i][0],MPIU_SCALAR,req_source2[i],tag4,comm,r_waits4+i);CHKERRQ(ierr); 2618 } 2619 2620 /* Allocate sending buffers for a->a, and send them off */ 2621 ierr = PetscMalloc1(nrqr+1,&sbuf_aa);CHKERRQ(ierr); 2622 for (i=0,j=0; i<nrqr; i++) j += req_size[i]; 2623 ierr = PetscMalloc1(j+1,&sbuf_aa[0]);CHKERRQ(ierr); 2624 for (i=1; i<nrqr; i++) sbuf_aa[i] = sbuf_aa[i-1] + req_size[i-1]; 2625 2626 ierr = PetscMalloc1(nrqr+1,&s_waits4);CHKERRQ(ierr); 2627 { 2628 PetscInt nzA,nzB,*a_i = a->i,*b_i = b->i, *cworkB,lwrite; 2629 PetscInt cstart = C->cmap->rstart,rstart = C->rmap->rstart,*bmap = c->garray; 2630 PetscInt cend = C->cmap->rend; 2631 PetscInt *b_j = b->j; 2632 PetscScalar *vworkA,*vworkB,*a_a = a->a,*b_a = b->a; 2633 2634 for (i=0; i<nrqr; i++) { 2635 rbuf1_i = rbuf1[i]; 2636 sbuf_aa_i = sbuf_aa[i]; 2637 ct1 = 2*rbuf1_i[0]+1; 2638 ct2 = 0; 2639 for (j=1,max1=rbuf1_i[0]; j<=max1; j++) { 2640 kmax = rbuf1_i[2*j]; 2641 for (k=0; k<kmax; k++,ct1++) { 2642 row = rbuf1_i[ct1] - rstart; 2643 nzA = a_i[row+1] - a_i[row]; nzB = b_i[row+1] - b_i[row]; 2644 ncols = nzA + nzB; 2645 cworkB = b_j + b_i[row]; 2646 vworkA = a_a + a_i[row]; 2647 vworkB = b_a + b_i[row]; 2648 2649 /* load the column values for this row into vals*/ 2650 vals = sbuf_aa_i+ct2; 2651 2652 lwrite = 0; 2653 for (l=0; l<nzB; l++) { 2654 if ((bmap[cworkB[l]]) < cstart) vals[lwrite++] = vworkB[l]; 2655 } 2656 for (l=0; l<nzA; l++) vals[lwrite++] = vworkA[l]; 2657 for (l=0; l<nzB; l++) { 2658 if ((bmap[cworkB[l]]) >= cend) vals[lwrite++] = vworkB[l]; 2659 } 2660 2661 ct2 += ncols; 2662 } 2663 } 2664 ierr = MPI_Isend(sbuf_aa_i,req_size[i],MPIU_SCALAR,req_source1[i],tag4,comm,s_waits4+i);CHKERRQ(ierr); 2665 } 2666 } 2667 2668 if (!ismax) { 2669 ierr = PetscFree(rbuf1[0]);CHKERRQ(ierr); 2670 ierr = PetscFree(rbuf1);CHKERRQ(ierr); 2671 } 2672 2673 /* Assemble the matrices */ 2674 /* First assemble the local rows */ 2675 { 2676 PetscInt ilen_row,*imat_ilen,*imat_j,*imat_i,old_row; 2677 PetscScalar *imat_a; 2678 2679 for (i=0; i<ismax; i++) { 2680 row2proc_i = row2proc[i]; 2681 subc = (Mat_SeqAIJ*)submats[i]->data; 2682 imat_ilen = subc->ilen; 2683 imat_j = subc->j; 2684 imat_i = subc->i; 2685 imat_a = subc->a; 2686 2687 if (!allcolumns[i]) cmap_i = cmap[i]; 2688 rmap_i = rmap[i]; 2689 irow_i = irow[i]; 2690 jmax = nrow[i]; 2691 for (j=0; j<jmax; j++) { 2692 row = irow_i[j]; 2693 proc = row2proc_i[j]; 2694 if (proc == rank) { 2695 old_row = row; 2696 #if defined(PETSC_USE_CTABLE) 2697 ierr = PetscTableFind(rmap_i,row+1,&row);CHKERRQ(ierr); 2698 row--; 2699 #else 2700 row = rmap_i[row]; 2701 #endif 2702 ilen_row = imat_ilen[row]; 2703 ierr = MatGetRow_MPIAIJ(C,old_row,&ncols,&cols,&vals);CHKERRQ(ierr); 2704 mat_i = imat_i[row]; 2705 mat_a = imat_a + mat_i; 2706 mat_j = imat_j + mat_i; 2707 if (!allcolumns[i]) { 2708 for (k=0; k<ncols; k++) { 2709 #if defined(PETSC_USE_CTABLE) 2710 ierr = PetscTableFind(cmap_i,cols[k]+1,&tcol);CHKERRQ(ierr); 2711 #else 2712 tcol = cmap_i[cols[k]]; 2713 #endif 2714 if (tcol) { 2715 *mat_j++ = tcol - 1; 2716 *mat_a++ = vals[k]; 2717 ilen_row++; 2718 } 2719 } 2720 } else { /* allcolumns */ 2721 for (k=0; k<ncols; k++) { 2722 *mat_j++ = cols[k]; /* global col index! */ 2723 *mat_a++ = vals[k]; 2724 ilen_row++; 2725 } 2726 } 2727 ierr = MatRestoreRow_MPIAIJ(C,old_row,&ncols,&cols,&vals);CHKERRQ(ierr); 2728 2729 imat_ilen[row] = ilen_row; 2730 } 2731 } 2732 } 2733 } 2734 2735 /* Now assemble the off proc rows*/ 2736 { 2737 PetscInt *sbuf1_i,*rbuf2_i,*rbuf3_i,*imat_ilen,ilen; 2738 PetscInt *imat_j,*imat_i; 2739 PetscScalar *imat_a,*rbuf4_i; 2740 2741 ierr = MPI_Waitall(nrqs,r_waits4,r_status4);CHKERRQ(ierr); 2742 for (tmp2=0; tmp2<nrqs; tmp2++) { 2743 sbuf1_i = sbuf1[pa[tmp2]]; 2744 jmax = sbuf1_i[0]; 2745 ct1 = 2*jmax + 1; 2746 ct2 = 0; 2747 rbuf2_i = rbuf2[tmp2]; 2748 rbuf3_i = rbuf3[tmp2]; 2749 rbuf4_i = rbuf4[tmp2]; 2750 for (j=1; j<=jmax; j++) { 2751 is_no = sbuf1_i[2*j-1]; 2752 rmap_i = rmap[is_no]; 2753 if (!allcolumns[is_no]) cmap_i = cmap[is_no]; 2754 subc = (Mat_SeqAIJ*)submats[is_no]->data; 2755 imat_ilen = subc->ilen; 2756 imat_j = subc->j; 2757 imat_i = subc->i; 2758 imat_a = subc->a; 2759 max1 = sbuf1_i[2*j]; 2760 for (k=0; k<max1; k++,ct1++) { 2761 row = sbuf1_i[ct1]; 2762 #if defined(PETSC_USE_CTABLE) 2763 ierr = PetscTableFind(rmap_i,row+1,&row);CHKERRQ(ierr); 2764 row--; 2765 #else 2766 row = rmap_i[row]; 2767 #endif 2768 ilen = imat_ilen[row]; 2769 mat_i = imat_i[row]; 2770 mat_a = imat_a + mat_i; 2771 mat_j = imat_j + mat_i; 2772 max2 = rbuf2_i[ct1]; 2773 if (!allcolumns[is_no]) { 2774 for (l=0; l<max2; l++,ct2++) { 2775 2776 #if defined(PETSC_USE_CTABLE) 2777 ierr = PetscTableFind(cmap_i,rbuf3_i[ct2]+1,&tcol);CHKERRQ(ierr); 2778 #else 2779 tcol = cmap_i[rbuf3_i[ct2]]; 2780 #endif 2781 if (tcol) { 2782 *mat_j++ = tcol - 1; 2783 *mat_a++ = rbuf4_i[ct2]; 2784 ilen++; 2785 } 2786 } 2787 } else { /* allcolumns */ 2788 for (l=0; l<max2; l++,ct2++) { 2789 *mat_j++ = rbuf3_i[ct2]; /* same global column index of C */ 2790 *mat_a++ = rbuf4_i[ct2]; 2791 ilen++; 2792 } 2793 } 2794 imat_ilen[row] = ilen; 2795 } 2796 } 2797 } 2798 } 2799 2800 ierr = PetscFree(r_status4);CHKERRQ(ierr); 2801 ierr = PetscFree(r_waits4);CHKERRQ(ierr); 2802 if (nrqr) {ierr = MPI_Waitall(nrqr,s_waits4,s_status4);CHKERRQ(ierr);} 2803 ierr = PetscFree(s_waits4);CHKERRQ(ierr); 2804 ierr = PetscFree(s_status4);CHKERRQ(ierr); 2805 2806 /* Restore the indices */ 2807 for (i=0; i<ismax; i++) { 2808 ierr = ISRestoreIndices(isrow[i],irow+i);CHKERRQ(ierr); 2809 if (!allcolumns[i]) { 2810 ierr = ISRestoreIndices(iscol[i],icol+i);CHKERRQ(ierr); 2811 } 2812 } 2813 2814 for (i=0; i<ismax; i++) { 2815 ierr = MatAssemblyBegin(submats[i],MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2816 ierr = MatAssemblyEnd(submats[i],MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2817 } 2818 2819 /* Destroy allocated memory */ 2820 if (!ismax) { 2821 ierr = PetscFree(pa);CHKERRQ(ierr); 2822 2823 ierr = PetscFree4(sbuf1,ptr,tmp,ctr);CHKERRQ(ierr); 2824 for (i=0; i<nrqr; ++i) { 2825 ierr = PetscFree(sbuf2[i]);CHKERRQ(ierr); 2826 } 2827 for (i=0; i<nrqs; ++i) { 2828 ierr = PetscFree(rbuf3[i]);CHKERRQ(ierr); 2829 } 2830 2831 ierr = PetscFree3(sbuf2,req_size,req_source1);CHKERRQ(ierr); 2832 ierr = PetscFree3(req_source2,rbuf2,rbuf3);CHKERRQ(ierr); 2833 } 2834 2835 ierr = PetscFree(sbuf_aa[0]);CHKERRQ(ierr); 2836 ierr = PetscFree(sbuf_aa);CHKERRQ(ierr); 2837 ierr = PetscFree5(irow,icol,nrow,ncol,issorted);CHKERRQ(ierr); 2838 2839 for (i=0; i<nrqs; ++i) { 2840 ierr = PetscFree(rbuf4[i]);CHKERRQ(ierr); 2841 } 2842 ierr = PetscFree(rbuf4);CHKERRQ(ierr); 2843 2844 ierr = PetscFree2(cmap,rmap);CHKERRQ(ierr); 2845 ierr = PetscFree(row2proc);CHKERRQ(ierr); 2846 ierr = PetscFree(smats);CHKERRQ(ierr); 2847 PetscFunctionReturn(0); 2848 } 2849 2850 /* 2851 Permute A & B into C's *local* index space using rowemb,dcolemb for A and rowemb,ocolemb for B. 2852 Embeddings are supposed to be injections and the above implies that the range of rowemb is a subset 2853 of [0,m), dcolemb is in [0,n) and ocolemb is in [N-n). 2854 If pattern == DIFFERENT_NONZERO_PATTERN, C is preallocated according to A&B. 2855 After that B's columns are mapped into C's global column space, so that C is in the "disassembled" 2856 state, and needs to be "assembled" later by compressing B's column space. 2857 2858 This function may be called in lieu of preallocation, so C should not be expected to be preallocated. 2859 Following this call, C->A & C->B have been created, even if empty. 2860 */ 2861 PetscErrorCode MatSetSeqMats_MPIAIJ(Mat C,IS rowemb,IS dcolemb,IS ocolemb,MatStructure pattern,Mat A,Mat B) 2862 { 2863 /* If making this function public, change the error returned in this function away from _PLIB. */ 2864 PetscErrorCode ierr; 2865 Mat_MPIAIJ *aij; 2866 Mat_SeqAIJ *Baij; 2867 PetscBool seqaij,Bdisassembled; 2868 PetscInt m,n,*nz,i,j,ngcol,col,rstart,rend,shift,count; 2869 PetscScalar v; 2870 const PetscInt *rowindices,*colindices; 2871 2872 PetscFunctionBegin; 2873 /* Check to make sure the component matrices (and embeddings) are compatible with C. */ 2874 if (A) { 2875 ierr = PetscObjectTypeCompare((PetscObject)A,MATSEQAIJ,&seqaij);CHKERRQ(ierr); 2876 if (!seqaij) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Diagonal matrix is of wrong type"); 2877 if (rowemb) { 2878 ierr = ISGetLocalSize(rowemb,&m);CHKERRQ(ierr); 2879 if (m != A->rmap->n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Row IS of size %D is incompatible with diag matrix row size %D",m,A->rmap->n); 2880 } else { 2881 if (C->rmap->n != A->rmap->n) { 2882 SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Diag seq matrix is row-incompatible with the MPIAIJ matrix"); 2883 } 2884 } 2885 if (dcolemb) { 2886 ierr = ISGetLocalSize(dcolemb,&n);CHKERRQ(ierr); 2887 if (n != A->cmap->n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Diag col IS of size %D is incompatible with diag matrix col size %D",n,A->cmap->n); 2888 } else { 2889 if (C->cmap->n != A->cmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Diag seq matrix is col-incompatible with the MPIAIJ matrix"); 2890 } 2891 } 2892 if (B) { 2893 ierr = PetscObjectTypeCompare((PetscObject)B,MATSEQAIJ,&seqaij);CHKERRQ(ierr); 2894 if (!seqaij) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Off-diagonal matrix is of wrong type"); 2895 if (rowemb) { 2896 ierr = ISGetLocalSize(rowemb,&m);CHKERRQ(ierr); 2897 if (m != B->rmap->n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Row IS of size %D is incompatible with off-diag matrix row size %D",m,A->rmap->n); 2898 } else { 2899 if (C->rmap->n != B->rmap->n) { 2900 SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Off-diag seq matrix is row-incompatible with the MPIAIJ matrix"); 2901 } 2902 } 2903 if (ocolemb) { 2904 ierr = ISGetLocalSize(ocolemb,&n);CHKERRQ(ierr); 2905 if (n != B->cmap->n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Off-diag col IS of size %D is incompatible with off-diag matrix col size %D",n,B->cmap->n); 2906 } else { 2907 if (C->cmap->N - C->cmap->n != B->cmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Off-diag seq matrix is col-incompatible with the MPIAIJ matrix"); 2908 } 2909 } 2910 2911 aij = (Mat_MPIAIJ*)(C->data); 2912 if (!aij->A) { 2913 /* Mimic parts of MatMPIAIJSetPreallocation() */ 2914 ierr = MatCreate(PETSC_COMM_SELF,&aij->A);CHKERRQ(ierr); 2915 ierr = MatSetSizes(aij->A,C->rmap->n,C->cmap->n,C->rmap->n,C->cmap->n);CHKERRQ(ierr); 2916 ierr = MatSetBlockSizesFromMats(aij->A,C,C);CHKERRQ(ierr); 2917 ierr = MatSetType(aij->A,MATSEQAIJ);CHKERRQ(ierr); 2918 ierr = PetscLogObjectParent((PetscObject)C,(PetscObject)aij->A);CHKERRQ(ierr); 2919 } 2920 if (A) { 2921 ierr = MatSetSeqMat_SeqAIJ(aij->A,rowemb,dcolemb,pattern,A);CHKERRQ(ierr); 2922 } else { 2923 ierr = MatSetUp(aij->A);CHKERRQ(ierr); 2924 } 2925 if (B) { /* Destroy the old matrix or the column map, depending on the sparsity pattern. */ 2926 /* 2927 If pattern == DIFFERENT_NONZERO_PATTERN, we reallocate B and 2928 need to "disassemble" B -- convert it to using C's global indices. 2929 To insert the values we take the safer, albeit more expensive, route of MatSetValues(). 2930 2931 If pattern == SUBSET_NONZERO_PATTERN, we do not "disassemble" B and do not reallocate; 2932 we MatZeroValues(B) first, so there may be a bunch of zeros that, perhaps, could be compacted out. 2933 2934 TODO: Put B's values into aij->B's aij structure in place using the embedding ISs? 2935 At least avoid calling MatSetValues() and the implied searches? 2936 */ 2937 2938 if (B && pattern == DIFFERENT_NONZERO_PATTERN) { 2939 #if defined(PETSC_USE_CTABLE) 2940 ierr = PetscTableDestroy(&aij->colmap);CHKERRQ(ierr); 2941 #else 2942 ierr = PetscFree(aij->colmap);CHKERRQ(ierr); 2943 /* A bit of a HACK: ideally we should deal with case aij->B all in one code block below. */ 2944 if (aij->B) { 2945 ierr = PetscLogObjectMemory((PetscObject)C,-aij->B->cmap->n*sizeof(PetscInt));CHKERRQ(ierr); 2946 } 2947 #endif 2948 ngcol = 0; 2949 if (aij->lvec) { 2950 ierr = VecGetSize(aij->lvec,&ngcol);CHKERRQ(ierr); 2951 } 2952 if (aij->garray) { 2953 ierr = PetscFree(aij->garray);CHKERRQ(ierr); 2954 ierr = PetscLogObjectMemory((PetscObject)C,-ngcol*sizeof(PetscInt));CHKERRQ(ierr); 2955 } 2956 ierr = VecDestroy(&aij->lvec);CHKERRQ(ierr); 2957 ierr = VecScatterDestroy(&aij->Mvctx);CHKERRQ(ierr); 2958 } 2959 if (aij->B && B && pattern == DIFFERENT_NONZERO_PATTERN) { 2960 ierr = MatDestroy(&aij->B);CHKERRQ(ierr); 2961 } 2962 if (aij->B && B && pattern == SUBSET_NONZERO_PATTERN) { 2963 ierr = MatZeroEntries(aij->B);CHKERRQ(ierr); 2964 } 2965 } 2966 Bdisassembled = PETSC_FALSE; 2967 if (!aij->B) { 2968 ierr = MatCreate(PETSC_COMM_SELF,&aij->B);CHKERRQ(ierr); 2969 ierr = PetscLogObjectParent((PetscObject)C,(PetscObject)aij->B);CHKERRQ(ierr); 2970 ierr = MatSetSizes(aij->B,C->rmap->n,C->cmap->N,C->rmap->n,C->cmap->N);CHKERRQ(ierr); 2971 ierr = MatSetBlockSizesFromMats(aij->B,B,B);CHKERRQ(ierr); 2972 ierr = MatSetType(aij->B,MATSEQAIJ);CHKERRQ(ierr); 2973 Bdisassembled = PETSC_TRUE; 2974 } 2975 if (B) { 2976 Baij = (Mat_SeqAIJ*)(B->data); 2977 if (pattern == DIFFERENT_NONZERO_PATTERN) { 2978 ierr = PetscMalloc1(B->rmap->n,&nz);CHKERRQ(ierr); 2979 for (i=0; i<B->rmap->n; i++) { 2980 nz[i] = Baij->i[i+1] - Baij->i[i]; 2981 } 2982 ierr = MatSeqAIJSetPreallocation(aij->B,0,nz);CHKERRQ(ierr); 2983 ierr = PetscFree(nz);CHKERRQ(ierr); 2984 } 2985 2986 ierr = PetscLayoutGetRange(C->rmap,&rstart,&rend);CHKERRQ(ierr); 2987 shift = rend-rstart; 2988 count = 0; 2989 rowindices = NULL; 2990 colindices = NULL; 2991 if (rowemb) { 2992 ierr = ISGetIndices(rowemb,&rowindices);CHKERRQ(ierr); 2993 } 2994 if (ocolemb) { 2995 ierr = ISGetIndices(ocolemb,&colindices);CHKERRQ(ierr); 2996 } 2997 for (i=0; i<B->rmap->n; i++) { 2998 PetscInt row; 2999 row = i; 3000 if (rowindices) row = rowindices[i]; 3001 for (j=Baij->i[i]; j<Baij->i[i+1]; j++) { 3002 col = Baij->j[count]; 3003 if (colindices) col = colindices[col]; 3004 if (Bdisassembled && col>=rstart) col += shift; 3005 v = Baij->a[count]; 3006 ierr = MatSetValues(aij->B,1,&row,1,&col,&v,INSERT_VALUES);CHKERRQ(ierr); 3007 ++count; 3008 } 3009 } 3010 /* No assembly for aij->B is necessary. */ 3011 /* FIXME: set aij->B's nonzerostate correctly. */ 3012 } else { 3013 ierr = MatSetUp(aij->B);CHKERRQ(ierr); 3014 } 3015 C->preallocated = PETSC_TRUE; 3016 C->was_assembled = PETSC_FALSE; 3017 C->assembled = PETSC_FALSE; 3018 /* 3019 C will need to be assembled so that aij->B can be compressed into local form in MatSetUpMultiply_MPIAIJ(). 3020 Furthermore, its nonzerostate will need to be based on that of aij->A's and aij->B's. 3021 */ 3022 PetscFunctionReturn(0); 3023 } 3024 3025 /* 3026 B uses local indices with column indices ranging between 0 and N-n; they must be interpreted using garray. 3027 */ 3028 PetscErrorCode MatGetSeqMats_MPIAIJ(Mat C,Mat *A,Mat *B) 3029 { 3030 Mat_MPIAIJ *aij = (Mat_MPIAIJ*) (C->data); 3031 3032 PetscFunctionBegin; 3033 PetscValidPointer(A,2); 3034 PetscValidPointer(B,3); 3035 /* FIXME: make sure C is assembled */ 3036 *A = aij->A; 3037 *B = aij->B; 3038 /* Note that we don't incref *A and *B, so be careful! */ 3039 PetscFunctionReturn(0); 3040 } 3041 3042 /* 3043 Extract MPI submatrices encoded by pairs of IS that may live on subcomms of C. 3044 NOT SCALABLE due to the use of ISGetNonlocalIS() (see below). 3045 */ 3046 PetscErrorCode MatGetSubMatricesMPI_MPIXAIJ(Mat C,PetscInt ismax,const IS isrow[],const IS iscol[],MatReuse scall,Mat *submat[], 3047 PetscErrorCode(*getsubmats_seq)(Mat,PetscInt,const IS[],const IS[],MatReuse,Mat**), 3048 PetscErrorCode(*getlocalmats)(Mat,Mat*,Mat*), 3049 PetscErrorCode(*setseqmat)(Mat,IS,IS,MatStructure,Mat), 3050 PetscErrorCode(*setseqmats)(Mat,IS,IS,IS,MatStructure,Mat,Mat)) 3051 { 3052 PetscErrorCode ierr; 3053 PetscMPIInt isize,flag; 3054 PetscInt i,ii,cismax,ispar; 3055 Mat *A,*B; 3056 IS *isrow_p,*iscol_p,*cisrow,*ciscol,*ciscol_p; 3057 3058 PetscFunctionBegin; 3059 if (!ismax) PetscFunctionReturn(0); 3060 3061 for (i = 0, cismax = 0; i < ismax; ++i) { 3062 PetscMPIInt isize; 3063 ierr = MPI_Comm_compare(((PetscObject)isrow[i])->comm,((PetscObject)iscol[i])->comm,&flag);CHKERRQ(ierr); 3064 if (flag != MPI_IDENT) SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Row and column index sets must have the same communicator"); 3065 ierr = MPI_Comm_size(((PetscObject)isrow[i])->comm, &isize);CHKERRQ(ierr); 3066 if (isize > 1) ++cismax; 3067 } 3068 3069 /* 3070 If cismax is zero on all C's ranks, then and only then can we use purely sequential matrix extraction. 3071 ispar counts the number of parallel ISs across C's comm. 3072 */ 3073 ierr = MPIU_Allreduce(&cismax,&ispar,1,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)C));CHKERRQ(ierr); 3074 if (!ispar) { /* Sequential ISs only across C's comm, so can call the sequential matrix extraction subroutine. */ 3075 ierr = (*getsubmats_seq)(C,ismax,isrow,iscol,scall,submat);CHKERRQ(ierr); 3076 PetscFunctionReturn(0); 3077 } 3078 3079 /* if (ispar) */ 3080 /* 3081 Construct the "complements" -- the off-processor indices -- of the iscol ISs for parallel ISs only. 3082 These are used to extract the off-diag portion of the resulting parallel matrix. 3083 The row IS for the off-diag portion is the same as for the diag portion, 3084 so we merely alias (without increfing) the row IS, while skipping those that are sequential. 3085 */ 3086 ierr = PetscMalloc2(cismax,&cisrow,cismax,&ciscol);CHKERRQ(ierr); 3087 ierr = PetscMalloc1(cismax,&ciscol_p);CHKERRQ(ierr); 3088 for (i = 0, ii = 0; i < ismax; ++i) { 3089 ierr = MPI_Comm_size(((PetscObject)isrow[i])->comm,&isize);CHKERRQ(ierr); 3090 if (isize > 1) { 3091 /* 3092 TODO: This is the part that's ***NOT SCALABLE***. 3093 To fix this we need to extract just the indices of C's nonzero columns 3094 that lie on the intersection of isrow[i] and ciscol[ii] -- the nonlocal 3095 part of iscol[i] -- without actually computing ciscol[ii]. This also has 3096 to be done without serializing on the IS list, so, most likely, it is best 3097 done by rewriting MatGetSubMatrices_MPIAIJ() directly. 3098 */ 3099 ierr = ISGetNonlocalIS(iscol[i],&(ciscol[ii]));CHKERRQ(ierr); 3100 /* Now we have to 3101 (a) make sure ciscol[ii] is sorted, since, even if the off-proc indices 3102 were sorted on each rank, concatenated they might no longer be sorted; 3103 (b) Use ISSortPermutation() to construct ciscol_p, the mapping from the 3104 indices in the nondecreasing order to the original index positions. 3105 If ciscol[ii] is strictly increasing, the permutation IS is NULL. 3106 */ 3107 ierr = ISSortPermutation(ciscol[ii],PETSC_FALSE,ciscol_p+ii);CHKERRQ(ierr); 3108 ierr = ISSort(ciscol[ii]);CHKERRQ(ierr); 3109 ++ii; 3110 } 3111 } 3112 ierr = PetscMalloc2(ismax,&isrow_p,ismax,&iscol_p);CHKERRQ(ierr); 3113 for (i = 0, ii = 0; i < ismax; ++i) { 3114 PetscInt j,issize; 3115 const PetscInt *indices; 3116 3117 /* 3118 Permute the indices into a nondecreasing order. Reject row and col indices with duplicates. 3119 */ 3120 ierr = ISSortPermutation(isrow[i],PETSC_FALSE,isrow_p+i);CHKERRQ(ierr); 3121 ierr = ISSort(isrow[i]);CHKERRQ(ierr); 3122 ierr = ISGetLocalSize(isrow[i],&issize);CHKERRQ(ierr); 3123 ierr = ISGetIndices(isrow[i],&indices);CHKERRQ(ierr); 3124 for (j = 1; j < issize; ++j) { 3125 if (indices[j] == indices[j-1]) { 3126 SETERRQ4(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Repeated indices in row IS %D: indices at %D and %D are both %D",i,j-1,j,indices[j]); 3127 } 3128 } 3129 ierr = ISRestoreIndices(isrow[i],&indices);CHKERRQ(ierr); 3130 3131 3132 ierr = ISSortPermutation(iscol[i],PETSC_FALSE,iscol_p+i);CHKERRQ(ierr); 3133 ierr = ISSort(iscol[i]);CHKERRQ(ierr); 3134 ierr = ISGetLocalSize(iscol[i],&issize);CHKERRQ(ierr); 3135 ierr = ISGetIndices(iscol[i],&indices);CHKERRQ(ierr); 3136 for (j = 1; j < issize; ++j) { 3137 if (indices[j-1] == indices[j]) { 3138 SETERRQ4(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Repeated indices in col IS %D: indices at %D and %D are both %D",i,j-1,j,indices[j]); 3139 } 3140 } 3141 ierr = ISRestoreIndices(iscol[i],&indices);CHKERRQ(ierr); 3142 ierr = MPI_Comm_size(((PetscObject)isrow[i])->comm,&isize);CHKERRQ(ierr); 3143 if (isize > 1) { 3144 cisrow[ii] = isrow[i]; 3145 ++ii; 3146 } 3147 } 3148 /* 3149 Allocate the necessary arrays to hold the resulting parallel matrices as well as the intermediate 3150 array of sequential matrices underlying the resulting parallel matrices. 3151 Which arrays to allocate is based on the value of MatReuse scall and whether ISs are sorted and/or 3152 contain duplicates. 3153 3154 There are as many diag matrices as there are original index sets. There are only as many parallel 3155 and off-diag matrices, as there are parallel (comm size > 1) index sets. 3156 3157 ARRAYS that can hold Seq matrices get allocated in any event -- either here or by getsubmats_seq(): 3158 - If the array of MPI matrices already exists and is being reused, we need to allocate the array 3159 and extract the underlying seq matrices into it to serve as placeholders, into which getsubmats_seq 3160 will deposite the extracted diag and off-diag parts. Thus, we allocate the A&B arrays and fill them 3161 with A[i] and B[ii] extracted from the corresponding MPI submat. 3162 - However, if the rows, A's column indices or B's column indices are not sorted, the extracted A[i] & B[ii] 3163 will have a different order from what getsubmats_seq expects. To handle this case -- indicated 3164 by a nonzero isrow_p[i], iscol_p[i], or ciscol_p[ii] -- we duplicate A[i] --> AA[i], B[ii] --> BB[ii] 3165 (retrieve composed AA[i] or BB[ii]) and reuse them here. AA[i] and BB[ii] are then used to permute its 3166 values into A[i] and B[ii] sitting inside the corresponding submat. 3167 - If no reuse is taking place then getsubmats_seq will allocate the A&B arrays and create the corresponding 3168 A[i], B[ii], AA[i] or BB[ii] matrices. 3169 */ 3170 /* Parallel matrix array is allocated here only if no reuse is taking place. If reused, it is passed in by the caller. */ 3171 if (scall == MAT_INITIAL_MATRIX) { 3172 ierr = PetscMalloc1(ismax,submat);CHKERRQ(ierr); 3173 } 3174 3175 /* Now obtain the sequential A and B submatrices separately. */ 3176 /* scall=MAT_REUSE_MATRIX is not handled yet, because getsubmats_seq() requires reuse of A and B */ 3177 ierr = (*getsubmats_seq)(C,ismax,isrow,iscol,MAT_INITIAL_MATRIX,&A);CHKERRQ(ierr); 3178 ierr = (*getsubmats_seq)(C,cismax,cisrow,ciscol,MAT_INITIAL_MATRIX,&B);CHKERRQ(ierr); 3179 3180 /* 3181 If scall == MAT_REUSE_MATRIX AND the permutations are NULL, we are done, since the sequential 3182 matrices A & B have been extracted directly into the parallel matrices containing them, or 3183 simply into the sequential matrix identical with the corresponding A (if isize == 1). 3184 Note that in that case colmap doesn't need to be rebuilt, since the matrices are expected 3185 to have the same sparsity pattern. 3186 Otherwise, A and/or B have to be properly embedded into C's index spaces and the correct colmap 3187 must be constructed for C. This is done by setseqmat(s). 3188 */ 3189 for (i = 0, ii = 0; i < ismax; ++i) { 3190 /* 3191 TODO: cache ciscol, permutation ISs and maybe cisrow? What about isrow & iscol? 3192 That way we can avoid sorting and computing permutations when reusing. 3193 To this end: 3194 - remove the old cache, if it exists, when extracting submatrices with MAT_INITIAL_MATRIX 3195 - if caching arrays to hold the ISs, make and compose a container for them so that it can 3196 be destroyed upon destruction of C (use PetscContainerUserDestroy() to clear out the contents). 3197 */ 3198 MatStructure pattern; 3199 pattern = DIFFERENT_NONZERO_PATTERN; 3200 3201 ierr = MPI_Comm_size(((PetscObject)isrow[i])->comm,&isize);CHKERRQ(ierr); 3202 /* Construct submat[i] from the Seq pieces A (and B, if necessary). */ 3203 if (isize > 1) { 3204 if (scall == MAT_INITIAL_MATRIX) { 3205 ierr = MatCreate(((PetscObject)isrow[i])->comm,(*submat)+i);CHKERRQ(ierr); 3206 ierr = MatSetSizes((*submat)[i],A[i]->rmap->n,A[i]->cmap->n,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 3207 ierr = MatSetType((*submat)[i],MATMPIAIJ);CHKERRQ(ierr); 3208 ierr = PetscLayoutSetUp((*submat)[i]->rmap);CHKERRQ(ierr); 3209 ierr = PetscLayoutSetUp((*submat)[i]->cmap);CHKERRQ(ierr); 3210 } 3211 /* 3212 For each parallel isrow[i], insert the extracted sequential matrices into the parallel matrix. 3213 */ 3214 { 3215 Mat AA,BB; 3216 AA = A[i]; 3217 BB = B[ii]; 3218 if (AA || BB) { 3219 ierr = setseqmats((*submat)[i],isrow_p[i],iscol_p[i],ciscol_p[ii],pattern,AA,BB);CHKERRQ(ierr); 3220 ierr = MatAssemblyBegin((*submat)[i],MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3221 ierr = MatAssemblyEnd((*submat)[i],MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3222 } 3223 3224 ierr = MatDestroy(&AA);CHKERRQ(ierr); 3225 ierr = MatDestroy(&BB);CHKERRQ(ierr); 3226 } 3227 ierr = ISDestroy(ciscol+ii);CHKERRQ(ierr); 3228 ierr = ISDestroy(ciscol_p+ii);CHKERRQ(ierr); 3229 ++ii; 3230 } else { /* if (isize == 1) */ 3231 if (scall == MAT_REUSE_MATRIX) { 3232 ierr = MatDestroy(&(*submat)[i]);CHKERRQ(ierr); 3233 } 3234 if (isrow_p[i] || iscol_p[i]) { 3235 ierr = MatDuplicate(A[i],MAT_DO_NOT_COPY_VALUES,(*submat)+i);CHKERRQ(ierr); 3236 ierr = setseqmat((*submat)[i],isrow_p[i],iscol_p[i],pattern,A[i]);CHKERRQ(ierr); 3237 /* Otherwise A is extracted straight into (*submats)[i]. */ 3238 /* TODO: Compose A[i] on (*submat([i] for future use, if ((isrow_p[i] || iscol_p[i]) && MAT_INITIAL_MATRIX). */ 3239 ierr = MatDestroy(A+i);CHKERRQ(ierr); 3240 } else (*submat)[i] = A[i]; 3241 } 3242 ierr = ISDestroy(&isrow_p[i]);CHKERRQ(ierr); 3243 ierr = ISDestroy(&iscol_p[i]);CHKERRQ(ierr); 3244 } 3245 ierr = PetscFree2(cisrow,ciscol);CHKERRQ(ierr); 3246 ierr = PetscFree2(isrow_p,iscol_p);CHKERRQ(ierr); 3247 ierr = PetscFree(ciscol_p);CHKERRQ(ierr); 3248 ierr = PetscFree(A);CHKERRQ(ierr); 3249 ierr = PetscFree(B);CHKERRQ(ierr); 3250 PetscFunctionReturn(0); 3251 } 3252 3253 PetscErrorCode MatGetSubMatricesMPI_MPIAIJ(Mat C,PetscInt ismax,const IS isrow[],const IS iscol[],MatReuse scall,Mat *submat[]) 3254 { 3255 PetscErrorCode ierr; 3256 3257 PetscFunctionBegin; 3258 ierr = MatGetSubMatricesMPI_MPIXAIJ(C,ismax,isrow,iscol,scall,submat,MatGetSubMatrices_MPIAIJ,MatGetSeqMats_MPIAIJ,MatSetSeqMat_SeqAIJ,MatSetSeqMats_MPIAIJ);CHKERRQ(ierr); 3259 PetscFunctionReturn(0); 3260 } 3261