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 PetscErrorCode MatDestroy_MPIAIJ_MatGetSubmatrices(Mat C) 1224 { 1225 PetscErrorCode ierr; 1226 Mat_SeqAIJ *c = (Mat_SeqAIJ*)C->data; 1227 Mat_SubMat *submatj = c->submatis1; 1228 PetscInt i; 1229 1230 PetscFunctionBegin; 1231 if (!submatj->id) { /* delete data that are linked only to submats[id=0] */ 1232 ierr = PetscFree4(submatj->sbuf1,submatj->ptr,submatj->tmp,submatj->ctr);CHKERRQ(ierr); 1233 1234 for (i=0; i<submatj->nrqr; ++i) { 1235 ierr = PetscFree(submatj->sbuf2[i]);CHKERRQ(ierr); 1236 } 1237 ierr = PetscFree3(submatj->sbuf2,submatj->req_size,submatj->req_source1);CHKERRQ(ierr); 1238 1239 if (submatj->rbuf1) { 1240 ierr = PetscFree(submatj->rbuf1[0]);CHKERRQ(ierr); 1241 ierr = PetscFree(submatj->rbuf1);CHKERRQ(ierr); 1242 } 1243 1244 for (i=0; i<submatj->nrqs; ++i) { 1245 ierr = PetscFree(submatj->rbuf3[i]);CHKERRQ(ierr); 1246 } 1247 ierr = PetscFree3(submatj->req_source2,submatj->rbuf2,submatj->rbuf3);CHKERRQ(ierr); 1248 ierr = PetscFree(submatj->pa);CHKERRQ(ierr); 1249 } 1250 1251 #if defined(PETSC_USE_CTABLE) 1252 ierr = PetscTableDestroy((PetscTable*)&submatj->rmap);CHKERRQ(ierr); 1253 if (submatj->cmap_loc) {ierr = PetscFree(submatj->cmap_loc);CHKERRQ(ierr);} 1254 ierr = PetscFree(submatj->rmap_loc);CHKERRQ(ierr); 1255 #else 1256 ierr = PetscFree(submatj->rmap);CHKERRQ(ierr); 1257 #endif 1258 1259 if (!submatj->allcolumns) { 1260 #if defined(PETSC_USE_CTABLE) 1261 ierr = PetscTableDestroy((PetscTable*)&submatj->cmap);CHKERRQ(ierr); 1262 #else 1263 ierr = PetscFree(submatj->cmap);CHKERRQ(ierr); 1264 #endif 1265 } 1266 ierr = submatj->destroy(C);CHKERRQ(ierr); 1267 ierr = PetscFree(submatj->row2proc);CHKERRQ(ierr); 1268 1269 ierr = PetscFree(submatj);CHKERRQ(ierr); 1270 PetscFunctionReturn(0); 1271 } 1272 1273 PetscErrorCode MatGetSubMatrices_MPIAIJ_SingleIS_Local(Mat C,PetscInt ismax,const IS isrow[],const IS iscol[],MatReuse scall,PetscBool allcolumns,Mat *submats) 1274 { 1275 Mat_MPIAIJ *c = (Mat_MPIAIJ*)C->data; 1276 Mat submat,A = c->A,B = c->B; 1277 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data,*b = (Mat_SeqAIJ*)B->data,*subc; 1278 PetscInt *ai=a->i,*aj=a->j,*bi=b->i,*bj=b->j,nzA,nzB; 1279 PetscInt cstart = C->cmap->rstart,cend = C->cmap->rend,rstart = C->rmap->rstart,*bmap = c->garray; 1280 const PetscInt *icol,*irow; 1281 PetscInt nrow,ncol,start; 1282 PetscErrorCode ierr; 1283 PetscMPIInt rank,size,tag1,tag2,tag3,tag4,*w1,*w2,nrqr; 1284 PetscInt **sbuf1,**sbuf2,i,j,k,l,ct1,ct2,ct3,**rbuf1,row,proc; 1285 PetscInt nrqs=0,msz,**ptr,*req_size,*ctr,*pa,*tmp,tcol,*iptr; 1286 PetscInt **rbuf3,*req_source1,*req_source2,**sbuf_aj,**rbuf2,max1,nnz; 1287 PetscInt *lens,rmax,ncols,*cols,Crow; 1288 #if defined(PETSC_USE_CTABLE) 1289 PetscTable cmap,rmap; 1290 PetscInt *cmap_loc,*rmap_loc; 1291 #else 1292 PetscInt *cmap,*rmap; 1293 #endif 1294 PetscInt ctr_j,*sbuf1_j,*sbuf_aj_i,*rbuf1_i,kmax,*sbuf1_i,*rbuf2_i,*rbuf3_i; 1295 PetscInt *cworkB,lwrite,*subcols,*row2proc; 1296 PetscScalar *vworkA,*vworkB,*a_a = a->a,*b_a = b->a,*subvals=NULL; 1297 MPI_Request *s_waits1,*r_waits1,*s_waits2,*r_waits2,*r_waits3; 1298 MPI_Request *r_waits4,*s_waits3 = NULL,*s_waits4; 1299 MPI_Status *r_status1,*r_status2,*s_status1,*s_status3 = NULL,*s_status2; 1300 MPI_Status *r_status3 = NULL,*r_status4,*s_status4; 1301 MPI_Comm comm; 1302 PetscScalar **rbuf4,**sbuf_aa,*vals,*sbuf_aa_i,*rbuf4_i; 1303 PetscMPIInt *onodes1,*olengths1,idex,end; 1304 Mat_SubMat *smatis1; 1305 PetscBool isrowsorted; 1306 1307 PetscFunctionBegin; 1308 if (ismax != 1) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"This routine only works when all processes have ismax=1"); 1309 1310 ierr = PetscObjectGetComm((PetscObject)C,&comm);CHKERRQ(ierr); 1311 size = c->size; 1312 rank = c->rank; 1313 1314 ierr = ISSorted(isrow[0],&isrowsorted);CHKERRQ(ierr); 1315 if (!isrowsorted) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"isrow[0] must be sorted"); 1316 1317 ierr = ISGetIndices(isrow[0],&irow);CHKERRQ(ierr); 1318 ierr = ISGetLocalSize(isrow[0],&nrow);CHKERRQ(ierr); 1319 if (allcolumns) { 1320 icol = NULL; 1321 ncol = C->cmap->N; 1322 } else { 1323 ierr = ISGetIndices(iscol[0],&icol);CHKERRQ(ierr); 1324 ierr = ISGetLocalSize(iscol[0],&ncol);CHKERRQ(ierr); 1325 } 1326 1327 if (scall == MAT_INITIAL_MATRIX) { 1328 PetscInt *sbuf2_i,*cworkA,lwrite,ctmp; 1329 1330 /* Get some new tags to keep the communication clean */ 1331 tag1 = ((PetscObject)C)->tag; 1332 ierr = PetscObjectGetNewTag((PetscObject)C,&tag2);CHKERRQ(ierr); 1333 ierr = PetscObjectGetNewTag((PetscObject)C,&tag3);CHKERRQ(ierr); 1334 1335 /* evaluate communication - mesg to who, length of mesg, and buffer space 1336 required. Based on this, buffers are allocated, and data copied into them */ 1337 ierr = PetscCalloc2(size,&w1,size,&w2);CHKERRQ(ierr); 1338 ierr = PetscMalloc1(nrow,&row2proc);CHKERRQ(ierr); 1339 1340 /* w1[proc] = num of rows owned by proc -- to be requested */ 1341 proc = 0; 1342 nrqs = 0; /* num of outgoing messages */ 1343 for (j=0; j<nrow; j++) { 1344 row = irow[j]; /* sorted! */ 1345 while (row >= C->rmap->range[proc+1]) proc++; 1346 w1[proc]++; 1347 row2proc[j] = proc; /* map row index to proc */ 1348 1349 if (proc != rank && !w2[proc]) { 1350 w2[proc] = 1; nrqs++; 1351 } 1352 } 1353 w1[rank] = 0; /* rows owned by self will not be requested */ 1354 1355 ierr = PetscMalloc1(nrqs+1,&pa);CHKERRQ(ierr); /*(proc -array)*/ 1356 for (proc=0,j=0; proc<size; proc++) { 1357 if (w1[proc]) { pa[j++] = proc;} 1358 } 1359 1360 /* Each message would have a header = 1 + 2*(num of IS) + data (here,num of IS = 1) */ 1361 msz = 0; /* total mesg length (for all procs) */ 1362 for (i=0; i<nrqs; i++) { 1363 proc = pa[i]; 1364 w1[proc] += 3; 1365 msz += w1[proc]; 1366 } 1367 ierr = PetscInfo2(0,"Number of outgoing messages %D Total message length %D\n",nrqs,msz);CHKERRQ(ierr); 1368 1369 /* Determine nrqr, the number of messages to expect, their lengths, from from-ids */ 1370 /* if w2[proc]=1, a message of length w1[proc] will be sent to proc; */ 1371 ierr = PetscGatherNumberOfMessages(comm,w2,w1,&nrqr);CHKERRQ(ierr); 1372 1373 /* Input: nrqs: nsend; nrqr: nrecv; w1: msg length to be sent; 1374 Output: onodes1: recv node-ids; olengths1: corresponding recv message length */ 1375 ierr = PetscGatherMessageLengths(comm,nrqs,nrqr,w1,&onodes1,&olengths1);CHKERRQ(ierr); 1376 1377 /* Now post the Irecvs corresponding to these messages */ 1378 ierr = PetscPostIrecvInt(comm,tag1,nrqr,onodes1,olengths1,&rbuf1,&r_waits1);CHKERRQ(ierr); 1379 1380 ierr = PetscFree(onodes1);CHKERRQ(ierr); 1381 ierr = PetscFree(olengths1);CHKERRQ(ierr); 1382 1383 /* Allocate Memory for outgoing messages */ 1384 ierr = PetscMalloc4(size,&sbuf1,size,&ptr,2*msz,&tmp,size,&ctr);CHKERRQ(ierr); 1385 ierr = PetscMemzero(sbuf1,size*sizeof(PetscInt*));CHKERRQ(ierr); 1386 ierr = PetscMemzero(ptr,size*sizeof(PetscInt*));CHKERRQ(ierr); 1387 1388 /* subf1[pa[0]] = tmp, subf1[pa[i]] = subf1[pa[i-1]] + w1[pa[i-1]] */ 1389 iptr = tmp; 1390 for (i=0; i<nrqs; i++) { 1391 proc = pa[i]; 1392 sbuf1[proc] = iptr; 1393 iptr += w1[proc]; 1394 } 1395 1396 /* Form the outgoing messages */ 1397 /* Initialize the header space */ 1398 for (i=0; i<nrqs; i++) { 1399 proc = pa[i]; 1400 ierr = PetscMemzero(sbuf1[proc],3*sizeof(PetscInt));CHKERRQ(ierr); 1401 ptr[proc] = sbuf1[proc] + 3; 1402 } 1403 1404 /* Parse the isrow and copy data into outbuf */ 1405 ierr = PetscMemzero(ctr,size*sizeof(PetscInt));CHKERRQ(ierr); 1406 for (j=0; j<nrow; j++) { /* parse the indices of each IS */ 1407 proc = row2proc[j]; 1408 if (proc != rank) { /* copy to the outgoing buf*/ 1409 *ptr[proc] = irow[j]; 1410 ctr[proc]++; ptr[proc]++; 1411 } 1412 } 1413 1414 /* Update the headers for the current IS */ 1415 for (j=0; j<size; j++) { /* Can Optimise this loop too */ 1416 if ((ctr_j = ctr[j])) { 1417 sbuf1_j = sbuf1[j]; 1418 k = ++sbuf1_j[0]; 1419 sbuf1_j[2*k] = ctr_j; 1420 sbuf1_j[2*k-1] = 0; 1421 } 1422 } 1423 1424 /* Now post the sends */ 1425 ierr = PetscMalloc1(nrqs+1,&s_waits1);CHKERRQ(ierr); 1426 for (i=0; i<nrqs; ++i) { 1427 proc = pa[i]; 1428 ierr = MPI_Isend(sbuf1[proc],w1[proc],MPIU_INT,proc,tag1,comm,s_waits1+i);CHKERRQ(ierr); 1429 } 1430 1431 /* Post Receives to capture the buffer size */ 1432 ierr = PetscMalloc4(nrqs+1,&r_status2,nrqr+1,&s_waits2,nrqs+1,&r_waits2,nrqr+1,&s_status2);CHKERRQ(ierr); 1433 ierr = PetscMalloc3(nrqs+1,&req_source2,nrqs+1,&rbuf2,nrqs+1,&rbuf3);CHKERRQ(ierr); 1434 1435 rbuf2[0] = tmp + msz; 1436 for (i=1; i<nrqs; ++i) rbuf2[i] = rbuf2[i-1] + w1[pa[i-1]]; 1437 1438 for (i=0; i<nrqs; ++i) { 1439 proc = pa[i]; 1440 ierr = MPI_Irecv(rbuf2[i],w1[proc],MPIU_INT,proc,tag2,comm,r_waits2+i);CHKERRQ(ierr); 1441 } 1442 1443 ierr = PetscFree2(w1,w2);CHKERRQ(ierr); 1444 1445 /* Send to other procs the buf size they should allocate */ 1446 /* Receive messages*/ 1447 ierr = PetscMalloc1(nrqr+1,&r_status1);CHKERRQ(ierr); 1448 ierr = PetscMalloc3(nrqr,&sbuf2,nrqr,&req_size,nrqr,&req_source1);CHKERRQ(ierr); 1449 1450 ierr = MPI_Waitall(nrqr,r_waits1,r_status1);CHKERRQ(ierr); 1451 for (i=0; i<nrqr; ++i) { 1452 req_size[i] = 0; 1453 rbuf1_i = rbuf1[i]; 1454 start = 2*rbuf1_i[0] + 1; 1455 ierr = MPI_Get_count(r_status1+i,MPIU_INT,&end);CHKERRQ(ierr); 1456 ierr = PetscMalloc1(end+1,&sbuf2[i]);CHKERRQ(ierr); 1457 sbuf2_i = sbuf2[i]; 1458 for (j=start; j<end; j++) { 1459 k = rbuf1_i[j] - rstart; 1460 ncols = ai[k+1] - ai[k] + bi[k+1] - bi[k]; 1461 sbuf2_i[j] = ncols; 1462 req_size[i] += ncols; 1463 } 1464 req_source1[i] = r_status1[i].MPI_SOURCE; 1465 1466 /* form the header */ 1467 sbuf2_i[0] = req_size[i]; 1468 for (j=1; j<start; j++) sbuf2_i[j] = rbuf1_i[j]; 1469 1470 ierr = MPI_Isend(sbuf2_i,end,MPIU_INT,req_source1[i],tag2,comm,s_waits2+i);CHKERRQ(ierr); 1471 } 1472 1473 ierr = PetscFree(r_status1);CHKERRQ(ierr); 1474 ierr = PetscFree(r_waits1);CHKERRQ(ierr); 1475 1476 /* rbuf2 is received, Post recv column indices a->j */ 1477 ierr = MPI_Waitall(nrqs,r_waits2,r_status2);CHKERRQ(ierr); 1478 1479 ierr = PetscMalloc4(nrqs+1,&r_waits3,nrqr+1,&s_waits3,nrqs+1,&r_status3,nrqr+1,&s_status3);CHKERRQ(ierr); 1480 for (i=0; i<nrqs; ++i) { 1481 ierr = PetscMalloc1(rbuf2[i][0]+1,&rbuf3[i]);CHKERRQ(ierr); 1482 req_source2[i] = r_status2[i].MPI_SOURCE; 1483 ierr = MPI_Irecv(rbuf3[i],rbuf2[i][0],MPIU_INT,req_source2[i],tag3,comm,r_waits3+i);CHKERRQ(ierr); 1484 } 1485 1486 /* Wait on sends1 and sends2 */ 1487 ierr = PetscMalloc1(nrqs+1,&s_status1);CHKERRQ(ierr); 1488 ierr = MPI_Waitall(nrqs,s_waits1,s_status1);CHKERRQ(ierr); 1489 ierr = PetscFree(s_waits1);CHKERRQ(ierr); 1490 ierr = PetscFree(s_status1);CHKERRQ(ierr); 1491 1492 ierr = MPI_Waitall(nrqr,s_waits2,s_status2);CHKERRQ(ierr); 1493 ierr = PetscFree4(r_status2,s_waits2,r_waits2,s_status2);CHKERRQ(ierr); 1494 1495 /* Now allocate sending buffers for a->j, and send them off */ 1496 ierr = PetscMalloc1(nrqr+1,&sbuf_aj);CHKERRQ(ierr); 1497 for (i=0,j=0; i<nrqr; i++) j += req_size[i]; 1498 ierr = PetscMalloc1(j+1,&sbuf_aj[0]);CHKERRQ(ierr); 1499 for (i=1; i<nrqr; i++) sbuf_aj[i] = sbuf_aj[i-1] + req_size[i-1]; 1500 1501 for (i=0; i<nrqr; i++) { /* for each requested message */ 1502 rbuf1_i = rbuf1[i]; 1503 sbuf_aj_i = sbuf_aj[i]; 1504 ct1 = 2*rbuf1_i[0] + 1; 1505 ct2 = 0; 1506 /* max1=rbuf1_i[0]; if (max1 != 1) SETERRQ1(PETSC_COMM_SELF,0,"max1 %d != 1",max1); */ 1507 1508 kmax = rbuf1[i][2]; 1509 for (k=0; k<kmax; k++,ct1++) { /* for each row */ 1510 row = rbuf1_i[ct1] - rstart; 1511 nzA = ai[row+1] - ai[row]; 1512 nzB = bi[row+1] - bi[row]; 1513 ncols = nzA + nzB; 1514 cworkA = aj + ai[row]; cworkB = bj + bi[row]; 1515 1516 /* load the column indices for this row into cols*/ 1517 cols = sbuf_aj_i + ct2; 1518 1519 lwrite = 0; 1520 for (l=0; l<nzB; l++) { 1521 if ((ctmp = bmap[cworkB[l]]) < cstart) cols[lwrite++] = ctmp; 1522 } 1523 for (l=0; l<nzA; l++) cols[lwrite++] = cstart + cworkA[l]; 1524 for (l=0; l<nzB; l++) { 1525 if ((ctmp = bmap[cworkB[l]]) >= cend) cols[lwrite++] = ctmp; 1526 } 1527 1528 ct2 += ncols; 1529 } 1530 ierr = MPI_Isend(sbuf_aj_i,req_size[i],MPIU_INT,req_source1[i],tag3,comm,s_waits3+i);CHKERRQ(ierr); 1531 } 1532 1533 /* create column map (cmap): global col of C -> local col of submat */ 1534 #if defined(PETSC_USE_CTABLE) 1535 if (!allcolumns) { 1536 ierr = PetscTableCreate(ncol+1,C->cmap->N+1,&cmap);CHKERRQ(ierr); 1537 ierr = PetscCalloc1(C->cmap->n,&cmap_loc);CHKERRQ(ierr); 1538 for (j=0; j<ncol; j++) { /* use array cmap_loc[] for local col indices */ 1539 if (icol[j] >= cstart && icol[j] <cend) { 1540 cmap_loc[icol[j] - cstart] = j+1; 1541 } else { /* use PetscTable for non-local col indices */ 1542 ierr = PetscTableAdd(cmap,icol[j]+1,j+1,INSERT_VALUES);CHKERRQ(ierr); 1543 } 1544 } 1545 } else { 1546 cmap = NULL; 1547 cmap_loc = NULL; 1548 } 1549 ierr = PetscCalloc1(C->rmap->n,&rmap_loc);CHKERRQ(ierr); 1550 #else 1551 if (!allcolumns) { 1552 ierr = PetscCalloc1(C->cmap->N,&cmap);CHKERRQ(ierr); 1553 for (j=0; j<ncol; j++) cmap[icol[j]] = j+1; 1554 } else { 1555 cmap = NULL; 1556 } 1557 #endif 1558 1559 /* Create lens for MatSeqAIJSetPreallocation() */ 1560 ierr = PetscCalloc1(nrow,&lens);CHKERRQ(ierr); 1561 1562 /* Compute lens from local part of C */ 1563 for (j=0; j<nrow; j++) { 1564 row = irow[j]; 1565 proc = row2proc[j]; 1566 if (proc == rank) { 1567 /* diagonal part A = c->A */ 1568 ncols = ai[row-rstart+1] - ai[row-rstart]; 1569 cols = aj + ai[row-rstart]; 1570 if (!allcolumns) { 1571 for (k=0; k<ncols; k++) { 1572 #if defined(PETSC_USE_CTABLE) 1573 tcol = cmap_loc[cols[k]]; 1574 #else 1575 tcol = cmap[cols[k]+cstart]; 1576 #endif 1577 if (tcol) lens[j]++; 1578 } 1579 } else { /* allcolumns */ 1580 lens[j] = ncols; 1581 } 1582 1583 /* off-diagonal part B = c->B */ 1584 ncols = bi[row-rstart+1] - bi[row-rstart]; 1585 cols = bj + bi[row-rstart]; 1586 if (!allcolumns) { 1587 for (k=0; k<ncols; k++) { 1588 #if defined(PETSC_USE_CTABLE) 1589 ierr = PetscTableFind(cmap,bmap[cols[k]]+1,&tcol);CHKERRQ(ierr); 1590 #else 1591 tcol = cmap[bmap[cols[k]]]; 1592 #endif 1593 if (tcol) lens[j]++; 1594 } 1595 } else { /* allcolumns */ 1596 lens[j] += ncols; 1597 } 1598 } 1599 } 1600 1601 /* Create row map (rmap): global row of C -> local row of submat */ 1602 #if defined(PETSC_USE_CTABLE) 1603 ierr = PetscTableCreate(nrow+1,C->rmap->N+1,&rmap);CHKERRQ(ierr); 1604 for (j=0; j<nrow; j++) { 1605 row = irow[j]; 1606 proc = row2proc[j]; 1607 if (proc == rank) { /* a local row */ 1608 rmap_loc[row - rstart] = j; 1609 } else { 1610 ierr = PetscTableAdd(rmap,irow[j]+1,j+1,INSERT_VALUES);CHKERRQ(ierr); 1611 } 1612 } 1613 #else 1614 ierr = PetscCalloc1(C->rmap->N,&rmap);CHKERRQ(ierr); 1615 for (j=0; j<nrow; j++) { 1616 rmap[irow[j]] = j; 1617 } 1618 #endif 1619 1620 /* Update lens from offproc data */ 1621 /* recv a->j is done */ 1622 ierr = MPI_Waitall(nrqs,r_waits3,r_status3);CHKERRQ(ierr); 1623 for (i=0; i<nrqs; i++) { 1624 proc = pa[i]; 1625 sbuf1_i = sbuf1[proc]; 1626 /* jmax = sbuf1_i[0]; if (jmax != 1) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"jmax !=1"); */ 1627 ct1 = 2 + 1; 1628 ct2 = 0; 1629 rbuf2_i = rbuf2[i]; /* received length of C->j */ 1630 rbuf3_i = rbuf3[i]; /* received C->j */ 1631 1632 /* is_no = sbuf1_i[2*j-1]; if (is_no != 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"is_no !=0"); */ 1633 max1 = sbuf1_i[2]; 1634 for (k=0; k<max1; k++,ct1++) { 1635 #if defined(PETSC_USE_CTABLE) 1636 ierr = PetscTableFind(rmap,sbuf1_i[ct1]+1,&row);CHKERRQ(ierr); 1637 row--; 1638 if (row < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"row not found in table"); 1639 #else 1640 row = rmap[sbuf1_i[ct1]]; /* the row index in submat */ 1641 #endif 1642 /* Now, store row index of submat in sbuf1_i[ct1] */ 1643 sbuf1_i[ct1] = row; 1644 1645 nnz = rbuf2_i[ct1]; 1646 if (!allcolumns) { 1647 for (l=0; l<nnz; l++,ct2++) { 1648 #if defined(PETSC_USE_CTABLE) 1649 if (rbuf3_i[ct2] >= cstart && rbuf3_i[ct2] <cend) { 1650 tcol = cmap_loc[rbuf3_i[ct2] - cstart]; 1651 } else { 1652 ierr = PetscTableFind(cmap,rbuf3_i[ct2]+1,&tcol);CHKERRQ(ierr); 1653 } 1654 #else 1655 tcol = cmap[rbuf3_i[ct2]]; /* column index in submat */ 1656 #endif 1657 if (tcol) lens[row]++; 1658 } 1659 } else { /* allcolumns */ 1660 lens[row] += nnz; 1661 } 1662 } 1663 } 1664 ierr = MPI_Waitall(nrqr,s_waits3,s_status3);CHKERRQ(ierr); 1665 ierr = PetscFree4(r_waits3,s_waits3,r_status3,s_status3);CHKERRQ(ierr); 1666 1667 /* Create the submatrices */ 1668 ierr = MatCreate(PETSC_COMM_SELF,&submat);CHKERRQ(ierr); 1669 ierr = MatSetSizes(submat,nrow,ncol,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 1670 1671 ierr = ISGetBlockSize(isrow[0],&i);CHKERRQ(ierr); 1672 ierr = ISGetBlockSize(iscol[0],&j);CHKERRQ(ierr); 1673 ierr = MatSetBlockSizes(submat,i,j);CHKERRQ(ierr); 1674 ierr = MatSetType(submat,((PetscObject)A)->type_name);CHKERRQ(ierr); 1675 ierr = MatSeqAIJSetPreallocation(submat,0,lens);CHKERRQ(ierr); 1676 1677 /* create struct Mat_SubMat and attached it to submat */ 1678 ierr = PetscNew(&smatis1);CHKERRQ(ierr); 1679 subc = (Mat_SeqAIJ*)submat->data; 1680 subc->submatis1 = smatis1; 1681 1682 smatis1->id = 0; 1683 smatis1->nrqs = nrqs; 1684 smatis1->nrqr = nrqr; 1685 smatis1->rbuf1 = rbuf1; 1686 smatis1->rbuf2 = rbuf2; 1687 smatis1->rbuf3 = rbuf3; 1688 smatis1->sbuf2 = sbuf2; 1689 smatis1->req_source2 = req_source2; 1690 1691 smatis1->sbuf1 = sbuf1; 1692 smatis1->ptr = ptr; 1693 smatis1->tmp = tmp; 1694 smatis1->ctr = ctr; 1695 1696 smatis1->pa = pa; 1697 smatis1->req_size = req_size; 1698 smatis1->req_source1 = req_source1; 1699 1700 smatis1->allcolumns = allcolumns; 1701 smatis1->row2proc = row2proc; 1702 smatis1->rmap = rmap; 1703 smatis1->cmap = cmap; 1704 #if defined(PETSC_USE_CTABLE) 1705 smatis1->rmap_loc = rmap_loc; 1706 smatis1->cmap_loc = cmap_loc; 1707 #endif 1708 1709 smatis1->destroy = submat->ops->destroy; 1710 submat->ops->destroy = MatDestroy_MPIAIJ_MatGetSubmatrices; 1711 submat->factortype = C->factortype; 1712 1713 /* compute rmax */ 1714 rmax = 0; 1715 for (i=0; i<nrow; i++) rmax = PetscMax(rmax,lens[i]); 1716 1717 } else { /* scall == MAT_REUSE_MATRIX */ 1718 submat = submats[0]; 1719 if (submat->rmap->n != nrow || submat->cmap->n != ncol) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Cannot reuse matrix. wrong size"); 1720 1721 subc = (Mat_SeqAIJ*)submat->data; 1722 rmax = subc->rmax; 1723 smatis1 = subc->submatis1; 1724 nrqs = smatis1->nrqs; 1725 nrqr = smatis1->nrqr; 1726 rbuf1 = smatis1->rbuf1; 1727 rbuf2 = smatis1->rbuf2; 1728 rbuf3 = smatis1->rbuf3; 1729 req_source2 = smatis1->req_source2; 1730 1731 sbuf1 = smatis1->sbuf1; 1732 sbuf2 = smatis1->sbuf2; 1733 ptr = smatis1->ptr; 1734 tmp = smatis1->tmp; 1735 ctr = smatis1->ctr; 1736 1737 pa = smatis1->pa; 1738 req_size = smatis1->req_size; 1739 req_source1 = smatis1->req_source1; 1740 1741 allcolumns = smatis1->allcolumns; 1742 row2proc = smatis1->row2proc; 1743 rmap = smatis1->rmap; 1744 cmap = smatis1->cmap; 1745 #if defined(PETSC_USE_CTABLE) 1746 rmap_loc = smatis1->rmap_loc; 1747 cmap_loc = smatis1->cmap_loc; 1748 #endif 1749 } 1750 1751 /* Post recv matrix values */ 1752 ierr = PetscMalloc3(nrqs+1,&rbuf4, rmax,&subcols, rmax,&subvals);CHKERRQ(ierr); 1753 ierr = PetscMalloc4(nrqs+1,&r_waits4,nrqr+1,&s_waits4,nrqs+1,&r_status4,nrqr+1,&s_status4);CHKERRQ(ierr); 1754 ierr = PetscObjectGetNewTag((PetscObject)C,&tag4);CHKERRQ(ierr); 1755 for (i=0; i<nrqs; ++i) { 1756 ierr = PetscMalloc1(rbuf2[i][0]+1,&rbuf4[i]);CHKERRQ(ierr); 1757 ierr = MPI_Irecv(rbuf4[i],rbuf2[i][0],MPIU_SCALAR,req_source2[i],tag4,comm,r_waits4+i);CHKERRQ(ierr); 1758 } 1759 1760 /* Allocate sending buffers for a->a, and send them off */ 1761 ierr = PetscMalloc1(nrqr+1,&sbuf_aa);CHKERRQ(ierr); 1762 for (i=0,j=0; i<nrqr; i++) j += req_size[i]; 1763 ierr = PetscMalloc1(j+1,&sbuf_aa[0]);CHKERRQ(ierr); 1764 for (i=1; i<nrqr; i++) sbuf_aa[i] = sbuf_aa[i-1] + req_size[i-1]; 1765 1766 for (i=0; i<nrqr; i++) { 1767 rbuf1_i = rbuf1[i]; 1768 sbuf_aa_i = sbuf_aa[i]; 1769 ct1 = 2*rbuf1_i[0]+1; 1770 ct2 = 0; 1771 /* max1=rbuf1_i[0]; if (max1 != 1) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"max1 !=1"); */ 1772 1773 kmax = rbuf1_i[2]; 1774 for (k=0; k<kmax; k++,ct1++) { 1775 row = rbuf1_i[ct1] - rstart; 1776 nzA = ai[row+1] - ai[row]; 1777 nzB = bi[row+1] - bi[row]; 1778 ncols = nzA + nzB; 1779 cworkB = bj + bi[row]; 1780 vworkA = a_a + ai[row]; 1781 vworkB = b_a + bi[row]; 1782 1783 /* load the column values for this row into vals*/ 1784 vals = sbuf_aa_i + ct2; 1785 1786 lwrite = 0; 1787 for (l=0; l<nzB; l++) { 1788 if ((bmap[cworkB[l]]) < cstart) vals[lwrite++] = vworkB[l]; 1789 } 1790 for (l=0; l<nzA; l++) vals[lwrite++] = vworkA[l]; 1791 for (l=0; l<nzB; l++) { 1792 if ((bmap[cworkB[l]]) >= cend) vals[lwrite++] = vworkB[l]; 1793 } 1794 1795 ct2 += ncols; 1796 } 1797 ierr = MPI_Isend(sbuf_aa_i,req_size[i],MPIU_SCALAR,req_source1[i],tag4,comm,s_waits4+i);CHKERRQ(ierr); 1798 } 1799 1800 /* Assemble submat */ 1801 /* First assemble the local rows */ 1802 for (j=0; j<nrow; j++) { 1803 row = irow[j]; 1804 proc = row2proc[j]; 1805 if (proc == rank) { 1806 Crow = row - rstart; /* local row index of C */ 1807 #if defined(PETSC_USE_CTABLE) 1808 row = rmap_loc[Crow]; /* row index of submat */ 1809 #else 1810 row = rmap[row]; 1811 #endif 1812 1813 if (allcolumns) { 1814 /* diagonal part A = c->A */ 1815 ncols = ai[Crow+1] - ai[Crow]; 1816 cols = aj + ai[Crow]; 1817 vals = a->a + ai[Crow]; 1818 i = 0; 1819 for (k=0; k<ncols; k++) { 1820 subcols[i] = cols[k] + cstart; 1821 subvals[i++] = vals[k]; 1822 } 1823 1824 /* off-diagonal part B = c->B */ 1825 ncols = bi[Crow+1] - bi[Crow]; 1826 cols = bj + bi[Crow]; 1827 vals = b->a + bi[Crow]; 1828 for (k=0; k<ncols; k++) { 1829 subcols[i] = bmap[cols[k]]; 1830 subvals[i++] = vals[k]; 1831 } 1832 1833 ierr = MatSetValues_SeqAIJ(submat,1,&row,i,subcols,subvals,INSERT_VALUES);CHKERRQ(ierr); 1834 1835 } else { /* !allcolumns */ 1836 #if defined(PETSC_USE_CTABLE) 1837 /* diagonal part A = c->A */ 1838 ncols = ai[Crow+1] - ai[Crow]; 1839 cols = aj + ai[Crow]; 1840 vals = a->a + ai[Crow]; 1841 i = 0; 1842 for (k=0; k<ncols; k++) { 1843 tcol = cmap_loc[cols[k]]; 1844 if (tcol) { 1845 subcols[i] = --tcol; 1846 subvals[i++] = vals[k]; 1847 } 1848 } 1849 1850 /* off-diagonal part B = c->B */ 1851 ncols = bi[Crow+1] - bi[Crow]; 1852 cols = bj + bi[Crow]; 1853 vals = b->a + bi[Crow]; 1854 for (k=0; k<ncols; k++) { 1855 ierr = PetscTableFind(cmap,bmap[cols[k]]+1,&tcol);CHKERRQ(ierr); 1856 if (tcol) { 1857 subcols[i] = --tcol; 1858 subvals[i++] = vals[k]; 1859 } 1860 } 1861 #else 1862 /* diagonal part A = c->A */ 1863 ncols = ai[Crow+1] - ai[Crow]; 1864 cols = aj + ai[Crow]; 1865 vals = a->a + ai[Crow]; 1866 i = 0; 1867 for (k=0; k<ncols; k++) { 1868 tcol = cmap[cols[k]+cstart]; 1869 if (tcol) { 1870 subcols[i] = --tcol; 1871 subvals[i++] = vals[k]; 1872 } 1873 } 1874 1875 /* off-diagonal part B = c->B */ 1876 ncols = bi[Crow+1] - bi[Crow]; 1877 cols = bj + bi[Crow]; 1878 vals = b->a + bi[Crow]; 1879 for (k=0; k<ncols; k++) { 1880 tcol = cmap[bmap[cols[k]]]; 1881 if (tcol) { 1882 subcols[i] = --tcol; 1883 subvals[i++] = vals[k]; 1884 } 1885 } 1886 #endif 1887 ierr = MatSetValues_SeqAIJ(submat,1,&row,i,subcols,subvals,INSERT_VALUES);CHKERRQ(ierr); 1888 } 1889 } 1890 } 1891 1892 /* Now assemble the off-proc rows */ 1893 for (i=0; i<nrqs; i++) { /* for each requested message */ 1894 /* recv values from other processes */ 1895 ierr = MPI_Waitany(nrqs,r_waits4,&idex,r_status4+i);CHKERRQ(ierr); 1896 proc = pa[idex]; 1897 sbuf1_i = sbuf1[proc]; 1898 /* jmax = sbuf1_i[0]; if (jmax != 1)SETERRQ1(PETSC_COMM_SELF,0,"jmax %d != 1",jmax); */ 1899 ct1 = 2 + 1; 1900 ct2 = 0; /* count of received C->j */ 1901 ct3 = 0; /* count of received C->j that will be inserted into submat */ 1902 rbuf2_i = rbuf2[idex]; /* int** received length of C->j from other processes */ 1903 rbuf3_i = rbuf3[idex]; /* int** received C->j from other processes */ 1904 rbuf4_i = rbuf4[idex]; /* scalar** received C->a from other processes */ 1905 1906 /* is_no = sbuf1_i[2*j-1]; if (is_no != 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"is_no !=0"); */ 1907 max1 = sbuf1_i[2]; /* num of rows */ 1908 for (k=0; k<max1; k++,ct1++) { /* for each recved row */ 1909 row = sbuf1_i[ct1]; /* row index of submat */ 1910 if (!allcolumns) { 1911 idex = 0; 1912 if (scall == MAT_INITIAL_MATRIX) { 1913 nnz = rbuf2_i[ct1]; /* num of C entries in this row */ 1914 for (l=0; l<nnz; l++,ct2++) { /* for each recved column */ 1915 #if defined(PETSC_USE_CTABLE) 1916 if (rbuf3_i[ct2] >= cstart && rbuf3_i[ct2] <cend) { 1917 tcol = cmap_loc[rbuf3_i[ct2] - cstart]; 1918 } else { 1919 ierr = PetscTableFind(cmap,rbuf3_i[ct2]+1,&tcol);CHKERRQ(ierr); 1920 } 1921 #else 1922 tcol = cmap[rbuf3_i[ct2]]; 1923 #endif 1924 if (tcol) { 1925 subcols[idex] = --tcol; 1926 subvals[idex++] = rbuf4_i[ct2]; 1927 1928 /* We receive an entire column of C, but a subset of it needs to be inserted into submat. 1929 For reuse, we replace received C->j with index that should be inserted to submat */ 1930 rbuf3_i[ct3++] = ct2; 1931 } 1932 } 1933 ierr = MatSetValues_SeqAIJ(submat,1,&row,idex,subcols,subvals,INSERT_VALUES);CHKERRQ(ierr); 1934 1935 } else { /* scall == MAT_REUSE_MATRIX */ 1936 submat = submats[0]; 1937 subc = (Mat_SeqAIJ*)submat->data; 1938 1939 nnz = subc->i[row+1] - subc->i[row]; /* num of submat entries in this row */ 1940 for (l=0; l<nnz; l++) { 1941 ct2 = rbuf3_i[ct3++]; /* index of rbuf4_i[] which needs to be inserted into submat */ 1942 subvals[idex++] = rbuf4_i[ct2]; 1943 } 1944 1945 bj = subc->j + subc->i[row]; 1946 ierr = MatSetValues_SeqAIJ(submat,1,&row,nnz,bj,subvals,INSERT_VALUES);CHKERRQ(ierr); 1947 } 1948 } else { /* allcolumns */ 1949 nnz = rbuf2_i[ct1]; /* num of C entries in this row */ 1950 ierr = MatSetValues_SeqAIJ(submat,1,&row,nnz,rbuf3_i+ct2,rbuf4_i+ct2,INSERT_VALUES);CHKERRQ(ierr); 1951 ct2 += nnz; 1952 } 1953 } 1954 } 1955 1956 /* sending a->a are done */ 1957 ierr = MPI_Waitall(nrqr,s_waits4,s_status4);CHKERRQ(ierr); 1958 ierr = PetscFree4(r_waits4,s_waits4,r_status4,s_status4);CHKERRQ(ierr); 1959 1960 ierr = MatAssemblyBegin(submat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1961 ierr = MatAssemblyEnd(submat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1962 submats[0] = submat; 1963 1964 /* Restore the indices */ 1965 ierr = ISRestoreIndices(isrow[0],&irow);CHKERRQ(ierr); 1966 if (!allcolumns) { 1967 ierr = ISRestoreIndices(iscol[0],&icol);CHKERRQ(ierr); 1968 } 1969 1970 /* Destroy allocated memory */ 1971 for (i=0; i<nrqs; ++i) { 1972 ierr = PetscFree3(rbuf4[i],subcols,subvals);CHKERRQ(ierr); 1973 } 1974 ierr = PetscFree3(rbuf4,subcols,subvals);CHKERRQ(ierr); 1975 ierr = PetscFree(sbuf_aa[0]);CHKERRQ(ierr); 1976 ierr = PetscFree(sbuf_aa);CHKERRQ(ierr); 1977 1978 if (scall == MAT_INITIAL_MATRIX) { 1979 ierr = PetscFree(lens);CHKERRQ(ierr); 1980 ierr = PetscFree(sbuf_aj[0]);CHKERRQ(ierr); 1981 ierr = PetscFree(sbuf_aj);CHKERRQ(ierr); 1982 } 1983 PetscFunctionReturn(0); 1984 } 1985 1986 PetscErrorCode MatGetSubMatrices_MPIAIJ_SingleIS(Mat C,PetscInt ismax,const IS isrow[],const IS iscol[],MatReuse scall,Mat *submat[]) 1987 { 1988 PetscErrorCode ierr; 1989 PetscInt ncol; 1990 PetscBool colflag,allcolumns=PETSC_FALSE; 1991 1992 PetscFunctionBegin; 1993 /* Allocate memory to hold all the submatrices */ 1994 if (scall == MAT_INITIAL_MATRIX) { 1995 ierr = PetscMalloc1(1,submat);CHKERRQ(ierr); 1996 } 1997 1998 /* Check for special case: each processor gets entire matrix columns */ 1999 ierr = ISIdentity(iscol[0],&colflag);CHKERRQ(ierr); 2000 ierr = ISGetLocalSize(iscol[0],&ncol);CHKERRQ(ierr); 2001 if (colflag && ncol == C->cmap->N) allcolumns = PETSC_TRUE; 2002 2003 ierr = MatGetSubMatrices_MPIAIJ_SingleIS_Local(C,ismax,isrow,iscol,scall,allcolumns,*submat);CHKERRQ(ierr); 2004 PetscFunctionReturn(0); 2005 } 2006 2007 PetscErrorCode MatGetSubMatrices_MPIAIJ(Mat C,PetscInt ismax,const IS isrow[],const IS iscol[],MatReuse scall,Mat *submat[]) 2008 { 2009 PetscErrorCode ierr; 2010 PetscInt nmax,nstages_local,nstages,i,pos,max_no,nrow,ncol; 2011 PetscBool rowflag,colflag,wantallmatrix=PETSC_FALSE,twantallmatrix,*allcolumns; 2012 2013 PetscFunctionBegin; 2014 #if 0 2015 /* Check for special case: each processor gets entire matrix */ 2016 if (C->submat_singleis) { /* flag is set in PCSetUp_ASM() to skip several MPIU_Allreduce() */ 2017 ierr = MatGetSubMatrices_MPIAIJ_SingleIS(C,ismax,isrow,iscol,scall,submat);CHKERRQ(ierr); 2018 PetscFunctionReturn(0); 2019 } 2020 #endif 2021 2022 if (ismax == 1 && C->rmap->N == C->cmap->N) { 2023 ierr = ISIdentity(*isrow,&rowflag);CHKERRQ(ierr); 2024 ierr = ISIdentity(*iscol,&colflag);CHKERRQ(ierr); 2025 ierr = ISGetLocalSize(*isrow,&nrow);CHKERRQ(ierr); 2026 ierr = ISGetLocalSize(*iscol,&ncol);CHKERRQ(ierr); 2027 if (rowflag && colflag && nrow == C->rmap->N && ncol == C->cmap->N) { 2028 wantallmatrix = PETSC_TRUE; 2029 2030 ierr = PetscOptionsGetBool(((PetscObject)C)->options,((PetscObject)C)->prefix,"-use_fast_submatrix",&wantallmatrix,NULL);CHKERRQ(ierr); 2031 } 2032 } 2033 ierr = MPIU_Allreduce(&wantallmatrix,&twantallmatrix,1,MPIU_BOOL,MPI_MIN,PetscObjectComm((PetscObject)C));CHKERRQ(ierr); 2034 if (twantallmatrix) { 2035 ierr = MatGetSubMatrix_MPIAIJ_All(C,MAT_GET_VALUES,scall,submat);CHKERRQ(ierr); 2036 PetscFunctionReturn(0); 2037 } 2038 2039 /* Allocate memory to hold all the submatrices */ 2040 if (scall == MAT_INITIAL_MATRIX) { 2041 ierr = PetscMalloc1(ismax+1,submat);CHKERRQ(ierr); 2042 } 2043 2044 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"); 2045 2046 /* Check for special case: each processor gets entire matrix columns */ 2047 ierr = PetscMalloc1(ismax+1,&allcolumns);CHKERRQ(ierr); 2048 for (i=0; i<ismax; i++) { 2049 ierr = ISIdentity(iscol[i],&colflag);CHKERRQ(ierr); 2050 ierr = ISGetLocalSize(iscol[i],&ncol);CHKERRQ(ierr); 2051 if (colflag && ncol == C->cmap->N) { 2052 allcolumns[i] = PETSC_TRUE; 2053 } else { 2054 allcolumns[i] = PETSC_FALSE; 2055 } 2056 } 2057 2058 /* Determine the number of stages through which submatrices are done */ 2059 nmax = 20*1000000 / (C->cmap->N * sizeof(PetscInt)); 2060 2061 /* 2062 Each stage will extract nmax submatrices. 2063 nmax is determined by the matrix column dimension. 2064 If the original matrix has 20M columns, only one submatrix per stage is allowed, etc. 2065 */ 2066 if (!nmax) nmax = 1; 2067 nstages_local = ismax/nmax + ((ismax % nmax) ? 1 : 0); 2068 2069 /* Make sure every processor loops through the nstages */ 2070 ierr = MPIU_Allreduce(&nstages_local,&nstages,1,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)C));CHKERRQ(ierr); 2071 2072 for (i=0,pos=0; i<nstages; i++) { 2073 if (pos+nmax <= ismax) max_no = nmax; 2074 else if (pos == ismax) max_no = 0; 2075 else max_no = ismax-pos; 2076 ierr = MatGetSubMatrices_MPIAIJ_Local(C,max_no,isrow+pos,iscol+pos,scall,allcolumns+pos,*submat+pos);CHKERRQ(ierr); 2077 pos += max_no; 2078 } 2079 2080 ierr = PetscFree(allcolumns);CHKERRQ(ierr); 2081 PetscFunctionReturn(0); 2082 } 2083 2084 /* -------------------------------------------------------------------------*/ 2085 PetscErrorCode MatGetSubMatrices_MPIAIJ_Local(Mat C,PetscInt ismax,const IS isrow[],const IS iscol[],MatReuse scall,PetscBool *allcolumns,Mat *submats) 2086 { 2087 Mat_MPIAIJ *c = (Mat_MPIAIJ*)C->data; 2088 Mat A = c->A; 2089 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data,*b = (Mat_SeqAIJ*)c->B->data,*subc; 2090 const PetscInt **icol,**irow; 2091 PetscInt *nrow,*ncol,start; 2092 PetscErrorCode ierr; 2093 PetscMPIInt rank,size,tag0,tag2,tag3,tag4,*w1,*w2,*w3,*w4,nrqr; 2094 PetscInt **sbuf1,**sbuf2,i,j,k,l,ct1,ct2,**rbuf1,row,proc; 2095 PetscInt nrqs,msz,**ptr,*req_size,*ctr,*pa,*tmp,tcol; 2096 PetscInt **rbuf3,*req_source1,*req_source2,**sbuf_aj,**rbuf2,max1,max2; 2097 PetscInt **lens,is_no,ncols,*cols,mat_i,*mat_j,tmp2,jmax; 2098 #if defined(PETSC_USE_CTABLE) 2099 PetscTable *cmap,cmap_i=NULL,*rmap,rmap_i; 2100 #else 2101 PetscInt **cmap,*cmap_i=NULL,**rmap,*rmap_i; 2102 #endif 2103 const PetscInt *irow_i; 2104 PetscInt ctr_j,*sbuf1_j,*sbuf_aj_i,*rbuf1_i,kmax,*lens_i; 2105 MPI_Request *s_waits1,*r_waits1,*s_waits2,*r_waits2,*r_waits3; 2106 MPI_Request *r_waits4,*s_waits3,*s_waits4; 2107 MPI_Status *r_status1,*r_status2,*s_status1,*s_status3,*s_status2; 2108 MPI_Status *r_status3,*r_status4,*s_status4; 2109 MPI_Comm comm; 2110 PetscScalar **rbuf4,**sbuf_aa,*vals,*mat_a,*sbuf_aa_i; 2111 PetscMPIInt *onodes1,*olengths1,end; 2112 PetscInt **row2proc,*row2proc_i,rmax; 2113 Mat_SubMat **smats,*smat_i; 2114 PetscBool *issorted; 2115 2116 PetscFunctionBegin; 2117 ierr = PetscObjectGetComm((PetscObject)C,&comm);CHKERRQ(ierr); 2118 size = c->size; 2119 rank = c->rank; 2120 //if (ismax == 0) printf("[%d] MatGetSubMatrices_MPIAIJ_Local, scall %d, ismax = 0!!! \n",rank,scall); 2121 //printf("[%d] MatGetSubMatrices_MPIAIJ_Local, scall %d\n",rank,scall); 2122 2123 ierr = PetscMalloc4(ismax,&irow,ismax,&icol,ismax,&nrow,ismax,&ncol);CHKERRQ(ierr); 2124 ierr = PetscMalloc1(ismax,&issorted);CHKERRQ(ierr); 2125 2126 for (i=0; i<ismax; i++) { 2127 ierr = ISSorted(isrow[i],&issorted[i]);CHKERRQ(ierr); 2128 2129 ierr = ISGetIndices(isrow[i],&irow[i]);CHKERRQ(ierr); 2130 ierr = ISGetLocalSize(isrow[i],&nrow[i]);CHKERRQ(ierr); 2131 if (allcolumns[i]) { 2132 icol[i] = NULL; 2133 ncol[i] = C->cmap->N; 2134 } else { 2135 ierr = ISGetIndices(iscol[i],&icol[i]);CHKERRQ(ierr); 2136 ierr = ISGetLocalSize(iscol[i],&ncol[i]);CHKERRQ(ierr); 2137 } 2138 } 2139 2140 ierr = PetscMalloc1(ismax,&smats);CHKERRQ(ierr); 2141 ierr = PetscMalloc1(ismax,&row2proc);CHKERRQ(ierr); 2142 ierr = PetscMalloc1(ismax,&cmap);CHKERRQ(ierr); 2143 ierr = PetscMalloc1(ismax,&rmap);CHKERRQ(ierr); 2144 2145 if (scall == MAT_REUSE_MATRIX) { 2146 /* Assumes new rows are same length as the old rows,hence bug! */ 2147 for (i=0; i<ismax; i++) { 2148 subc = (Mat_SeqAIJ*)(submats[i]->data); 2149 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"); 2150 2151 /* Initial matrix as if empty -- rm later!!! */ 2152 ierr = PetscMemzero(subc->ilen,submats[i]->rmap->n*sizeof(PetscInt));CHKERRQ(ierr); 2153 2154 /* Initial matrix as if empty */ 2155 submats[i]->factortype = C->factortype; 2156 2157 smat_i = subc->submatis1; 2158 smats[i] = smat_i; 2159 2160 nrqs = smat_i->nrqs; 2161 nrqr = smat_i->nrqr; 2162 rbuf1 = smat_i->rbuf1; 2163 rbuf2 = smat_i->rbuf2; 2164 rbuf3 = smat_i->rbuf3; 2165 req_source2 = smat_i->req_source2; 2166 2167 sbuf1 = smat_i->sbuf1; 2168 sbuf2 = smat_i->sbuf2; 2169 ptr = smat_i->ptr; 2170 tmp = smat_i->tmp; 2171 ctr = smat_i->ctr; 2172 2173 pa = smat_i->pa; 2174 req_size = smat_i->req_size; 2175 req_source1 = smat_i->req_source1; 2176 2177 allcolumns[i] = smat_i->allcolumns; 2178 row2proc[i] = smat_i->row2proc; 2179 rmap[i] = smat_i->rmap; 2180 cmap[i] = smat_i->cmap; 2181 #if defined(PETSC_USE_CTABLE) 2182 //rmap_loc = smat_i->rmap_loc; 2183 //cmap_loc = smat_i->cmap_loc; 2184 #endif 2185 } 2186 } else { /* scall == MAT_INITIAL_MATRIX */ 2187 /* Get some new tags to keep the communication clean */ 2188 ierr = PetscObjectGetNewTag((PetscObject)C,&tag2);CHKERRQ(ierr); 2189 ierr = PetscObjectGetNewTag((PetscObject)C,&tag3);CHKERRQ(ierr); 2190 2191 /* evaluate communication - mesg to who, length of mesg, and buffer space 2192 required. Based on this, buffers are allocated, and data copied into them*/ 2193 ierr = PetscMalloc4(size,&w1,size,&w2,size,&w3,size,&w4);CHKERRQ(ierr); /* mesg size */ 2194 ierr = PetscMemzero(w1,size*sizeof(PetscMPIInt));CHKERRQ(ierr); /* initialize work vector*/ 2195 ierr = PetscMemzero(w2,size*sizeof(PetscMPIInt));CHKERRQ(ierr); /* initialize work vector*/ 2196 ierr = PetscMemzero(w3,size*sizeof(PetscMPIInt));CHKERRQ(ierr); /* initialize work vector*/ 2197 2198 for (i=0; i<ismax; i++) { 2199 ierr = PetscMemzero(w4,size*sizeof(PetscMPIInt));CHKERRQ(ierr); /* initialize work vector*/ 2200 jmax = nrow[i]; 2201 irow_i = irow[i]; 2202 2203 ierr = PetscMalloc1(jmax,&row2proc_i);CHKERRQ(ierr); 2204 row2proc[i] = row2proc_i; 2205 2206 if (issorted[i]) proc = 0; 2207 for (j=0; j<jmax; j++) { 2208 if (!issorted[i]) proc = 0; 2209 row = irow_i[j]; 2210 while (row >= C->rmap->range[proc+1]) proc++; 2211 w4[proc]++; 2212 row2proc_i[j] = proc; /* map row index to proc */ 2213 } 2214 for (j=0; j<size; j++) { 2215 if (w4[j]) { w1[j] += w4[j]; w3[j]++;} 2216 } 2217 } 2218 2219 nrqs = 0; /* no of outgoing messages */ 2220 msz = 0; /* total mesg length (for all procs) */ 2221 w1[rank] = 0; /* no mesg sent to self */ 2222 w3[rank] = 0; 2223 for (i=0; i<size; i++) { 2224 if (w1[i]) { w2[i] = 1; nrqs++;} /* there exists a message to proc i */ 2225 } 2226 ierr = PetscMalloc1(nrqs+1,&pa);CHKERRQ(ierr); /*(proc -array)*/ 2227 for (i=0,j=0; i<size; i++) { 2228 if (w1[i]) { pa[j] = i; j++; } 2229 } 2230 2231 /* Each message would have a header = 1 + 2*(no of IS) + data */ 2232 for (i=0; i<nrqs; i++) { 2233 j = pa[i]; 2234 w1[j] += w2[j] + 2* w3[j]; 2235 msz += w1[j]; 2236 } 2237 ierr = PetscInfo2(0,"Number of outgoing messages %D Total message length %D\n",nrqs,msz);CHKERRQ(ierr); 2238 2239 /* Determine the number of messages to expect, their lengths, from from-ids */ 2240 ierr = PetscGatherNumberOfMessages(comm,w2,w1,&nrqr);CHKERRQ(ierr); 2241 ierr = PetscGatherMessageLengths(comm,nrqs,nrqr,w1,&onodes1,&olengths1);CHKERRQ(ierr); 2242 2243 /* Now post the Irecvs corresponding to these messages */ 2244 tag0 = ((PetscObject)C)->tag; 2245 ierr = PetscPostIrecvInt(comm,tag0,nrqr,onodes1,olengths1,&rbuf1,&r_waits1);CHKERRQ(ierr); 2246 2247 ierr = PetscFree(onodes1);CHKERRQ(ierr); 2248 ierr = PetscFree(olengths1);CHKERRQ(ierr); 2249 2250 /* Allocate Memory for outgoing messages */ 2251 ierr = PetscMalloc4(size,&sbuf1,size,&ptr,2*msz,&tmp,size,&ctr);CHKERRQ(ierr); 2252 ierr = PetscMemzero(sbuf1,size*sizeof(PetscInt*));CHKERRQ(ierr); 2253 ierr = PetscMemzero(ptr,size*sizeof(PetscInt*));CHKERRQ(ierr); 2254 2255 { 2256 PetscInt *iptr = tmp; 2257 k = 0; 2258 for (i=0; i<nrqs; i++) { 2259 j = pa[i]; 2260 iptr += k; 2261 sbuf1[j] = iptr; 2262 k = w1[j]; 2263 } 2264 } 2265 2266 /* Form the outgoing messages */ 2267 /* Initialize the header space */ 2268 for (i=0; i<nrqs; i++) { 2269 j = pa[i]; 2270 sbuf1[j][0] = 0; 2271 ierr = PetscMemzero(sbuf1[j]+1,2*w3[j]*sizeof(PetscInt));CHKERRQ(ierr); 2272 ptr[j] = sbuf1[j] + 2*w3[j] + 1; 2273 } 2274 2275 /* Parse the isrow and copy data into outbuf */ 2276 for (i=0; i<ismax; i++) { 2277 row2proc_i = row2proc[i]; 2278 ierr = PetscMemzero(ctr,size*sizeof(PetscInt));CHKERRQ(ierr); 2279 irow_i = irow[i]; 2280 jmax = nrow[i]; 2281 for (j=0; j<jmax; j++) { /* parse the indices of each IS */ 2282 proc = row2proc_i[j]; 2283 if (proc != rank) { /* copy to the outgoing buf*/ 2284 ctr[proc]++; 2285 *ptr[proc] = irow_i[j]; 2286 ptr[proc]++; 2287 } 2288 } 2289 /* Update the headers for the current IS */ 2290 for (j=0; j<size; j++) { /* Can Optimise this loop too */ 2291 if ((ctr_j = ctr[j])) { 2292 sbuf1_j = sbuf1[j]; 2293 k = ++sbuf1_j[0]; 2294 sbuf1_j[2*k] = ctr_j; 2295 sbuf1_j[2*k-1] = i; 2296 } 2297 } 2298 } 2299 2300 /* Now post the sends */ 2301 ierr = PetscMalloc1(nrqs+1,&s_waits1);CHKERRQ(ierr); 2302 for (i=0; i<nrqs; ++i) { 2303 j = pa[i]; 2304 ierr = MPI_Isend(sbuf1[j],w1[j],MPIU_INT,j,tag0,comm,s_waits1+i);CHKERRQ(ierr); 2305 } 2306 2307 /* Post Receives to capture the buffer size */ 2308 ierr = PetscMalloc1(nrqs+1,&r_waits2);CHKERRQ(ierr); 2309 ierr = PetscMalloc3(nrqs+1,&req_source2,nrqs+1,&rbuf2,nrqs+1,&rbuf3);CHKERRQ(ierr); 2310 rbuf2[0] = tmp + msz; 2311 for (i=1; i<nrqs; ++i) { 2312 rbuf2[i] = rbuf2[i-1]+w1[pa[i-1]]; 2313 } 2314 for (i=0; i<nrqs; ++i) { 2315 j = pa[i]; 2316 ierr = MPI_Irecv(rbuf2[i],w1[j],MPIU_INT,j,tag2,comm,r_waits2+i);CHKERRQ(ierr); 2317 } 2318 2319 /* Send to other procs the buf size they should allocate */ 2320 /* Receive messages*/ 2321 ierr = PetscMalloc1(nrqr+1,&s_waits2);CHKERRQ(ierr); 2322 ierr = PetscMalloc1(nrqr+1,&r_status1);CHKERRQ(ierr); 2323 ierr = PetscMalloc3(nrqr,&sbuf2,nrqr,&req_size,nrqr,&req_source1);CHKERRQ(ierr); 2324 { 2325 PetscInt *sAi = a->i,*sBi = b->i,id,rstart = C->rmap->rstart; 2326 PetscInt *sbuf2_i; 2327 2328 ierr = MPI_Waitall(nrqr,r_waits1,r_status1);CHKERRQ(ierr); 2329 for (i=0; i<nrqr; ++i) { 2330 req_size[i] = 0; 2331 rbuf1_i = rbuf1[i]; 2332 start = 2*rbuf1_i[0] + 1; 2333 ierr = MPI_Get_count(r_status1+i,MPIU_INT,&end);CHKERRQ(ierr); 2334 ierr = PetscMalloc1(end+1,&sbuf2[i]);CHKERRQ(ierr); 2335 sbuf2_i = sbuf2[i]; 2336 for (j=start; j<end; j++) { 2337 id = rbuf1_i[j] - rstart; 2338 ncols = sAi[id+1] - sAi[id] + sBi[id+1] - sBi[id]; 2339 sbuf2_i[j] = ncols; 2340 req_size[i] += ncols; 2341 } 2342 req_source1[i] = r_status1[i].MPI_SOURCE; 2343 /* form the header */ 2344 sbuf2_i[0] = req_size[i]; 2345 for (j=1; j<start; j++) sbuf2_i[j] = rbuf1_i[j]; 2346 2347 ierr = MPI_Isend(sbuf2_i,end,MPIU_INT,req_source1[i],tag2,comm,s_waits2+i);CHKERRQ(ierr); 2348 } 2349 } 2350 ierr = PetscFree(r_status1);CHKERRQ(ierr); 2351 ierr = PetscFree(r_waits1);CHKERRQ(ierr); 2352 2353 /* Receive messages*/ 2354 ierr = PetscMalloc1(nrqs+1,&r_waits3);CHKERRQ(ierr); 2355 ierr = PetscMalloc1(nrqs+1,&r_status2);CHKERRQ(ierr); 2356 2357 ierr = MPI_Waitall(nrqs,r_waits2,r_status2);CHKERRQ(ierr); 2358 for (i=0; i<nrqs; ++i) { 2359 ierr = PetscMalloc1(rbuf2[i][0]+1,&rbuf3[i]);CHKERRQ(ierr); 2360 req_source2[i] = r_status2[i].MPI_SOURCE; 2361 ierr = MPI_Irecv(rbuf3[i],rbuf2[i][0],MPIU_INT,req_source2[i],tag3,comm,r_waits3+i);CHKERRQ(ierr); 2362 } 2363 ierr = PetscFree(r_status2);CHKERRQ(ierr); 2364 ierr = PetscFree(r_waits2);CHKERRQ(ierr); 2365 2366 /* Wait on sends1 and sends2 */ 2367 ierr = PetscMalloc1(nrqs+1,&s_status1);CHKERRQ(ierr); 2368 ierr = PetscMalloc1(nrqr+1,&s_status2);CHKERRQ(ierr); 2369 2370 if (nrqs) {ierr = MPI_Waitall(nrqs,s_waits1,s_status1);CHKERRQ(ierr);} 2371 if (nrqr) {ierr = MPI_Waitall(nrqr,s_waits2,s_status2);CHKERRQ(ierr);} 2372 ierr = PetscFree(s_status1);CHKERRQ(ierr); 2373 ierr = PetscFree(s_status2);CHKERRQ(ierr); 2374 ierr = PetscFree(s_waits1);CHKERRQ(ierr); 2375 ierr = PetscFree(s_waits2);CHKERRQ(ierr); 2376 2377 /* Now allocate sending buffers for a->j, and send them off */ 2378 ierr = PetscMalloc1(nrqr+1,&sbuf_aj);CHKERRQ(ierr); 2379 for (i=0,j=0; i<nrqr; i++) j += req_size[i]; 2380 ierr = PetscMalloc1(j+1,&sbuf_aj[0]);CHKERRQ(ierr); 2381 for (i=1; i<nrqr; i++) sbuf_aj[i] = sbuf_aj[i-1] + req_size[i-1]; 2382 2383 ierr = PetscMalloc1(nrqr+1,&s_waits3);CHKERRQ(ierr); 2384 { 2385 PetscInt nzA,nzB,*a_i = a->i,*b_i = b->i,lwrite; 2386 PetscInt *cworkA,*cworkB,cstart = C->cmap->rstart,rstart = C->rmap->rstart,*bmap = c->garray; 2387 PetscInt cend = C->cmap->rend; 2388 PetscInt *a_j = a->j,*b_j = b->j,ctmp; 2389 2390 for (i=0; i<nrqr; i++) { 2391 rbuf1_i = rbuf1[i]; 2392 sbuf_aj_i = sbuf_aj[i]; 2393 ct1 = 2*rbuf1_i[0] + 1; 2394 ct2 = 0; 2395 for (j=1,max1=rbuf1_i[0]; j<=max1; j++) { 2396 kmax = rbuf1[i][2*j]; 2397 for (k=0; k<kmax; k++,ct1++) { 2398 row = rbuf1_i[ct1] - rstart; 2399 nzA = a_i[row+1] - a_i[row]; nzB = b_i[row+1] - b_i[row]; 2400 ncols = nzA + nzB; 2401 cworkA = a_j + a_i[row]; cworkB = b_j + b_i[row]; 2402 2403 /* load the column indices for this row into cols*/ 2404 cols = sbuf_aj_i + ct2; 2405 2406 lwrite = 0; 2407 for (l=0; l<nzB; l++) { 2408 if ((ctmp = bmap[cworkB[l]]) < cstart) cols[lwrite++] = ctmp; 2409 } 2410 for (l=0; l<nzA; l++) cols[lwrite++] = cstart + cworkA[l]; 2411 for (l=0; l<nzB; l++) { 2412 if ((ctmp = bmap[cworkB[l]]) >= cend) cols[lwrite++] = ctmp; 2413 } 2414 2415 ct2 += ncols; 2416 } 2417 } 2418 ierr = MPI_Isend(sbuf_aj_i,req_size[i],MPIU_INT,req_source1[i],tag3,comm,s_waits3+i);CHKERRQ(ierr); 2419 } 2420 } 2421 ierr = PetscMalloc1(nrqs+1,&r_status3);CHKERRQ(ierr); 2422 ierr = PetscMalloc1(nrqr+1,&s_status3);CHKERRQ(ierr); 2423 2424 /* create col map: global col of C -> local col of submatrices */ 2425 { 2426 const PetscInt *icol_i; 2427 #if defined(PETSC_USE_CTABLE) 2428 for (i=0; i<ismax; i++) { 2429 if (!allcolumns[i]) { 2430 ierr = PetscTableCreate(ncol[i]+1,C->cmap->N+1,&cmap[i]);CHKERRQ(ierr); 2431 2432 jmax = ncol[i]; 2433 icol_i = icol[i]; 2434 cmap_i = cmap[i]; 2435 for (j=0; j<jmax; j++) { 2436 ierr = PetscTableAdd(cmap[i],icol_i[j]+1,j+1,INSERT_VALUES);CHKERRQ(ierr); 2437 } 2438 } else cmap[i] = NULL; 2439 } 2440 #else 2441 for (i=0; i<ismax; i++) { 2442 if (!allcolumns[i]) { 2443 ierr = PetscMalloc1(C->cmap->N,&cmap[i]);CHKERRQ(ierr); 2444 ierr = PetscMemzero(cmap[i],C->cmap->N*sizeof(PetscInt));CHKERRQ(ierr); 2445 jmax = ncol[i]; 2446 icol_i = icol[i]; 2447 cmap_i = cmap[i]; 2448 for (j=0; j<jmax; j++) { 2449 cmap_i[icol_i[j]] = j+1; 2450 } 2451 } else cmap[i] = NULL; 2452 } 2453 #endif 2454 } 2455 2456 /* Create lens which is required for MatCreate... */ 2457 for (i=0,j=0; i<ismax; i++) j += nrow[i]; 2458 ierr = PetscMalloc1(ismax,&lens);CHKERRQ(ierr); 2459 2460 if (ismax) { 2461 ierr = PetscMalloc1(j,&lens[0]);CHKERRQ(ierr); 2462 ierr = PetscMemzero(lens[0],j*sizeof(PetscInt));CHKERRQ(ierr); 2463 } 2464 for (i=1; i<ismax; i++) lens[i] = lens[i-1] + nrow[i-1]; 2465 2466 /* Update lens from local data */ 2467 for (i=0; i<ismax; i++) { 2468 row2proc_i = row2proc[i]; 2469 jmax = nrow[i]; 2470 if (!allcolumns[i]) cmap_i = cmap[i]; 2471 irow_i = irow[i]; 2472 lens_i = lens[i]; 2473 for (j=0; j<jmax; j++) { 2474 row = irow_i[j]; 2475 proc = row2proc_i[j]; 2476 if (proc == rank) { 2477 ierr = MatGetRow_MPIAIJ(C,row,&ncols,&cols,0);CHKERRQ(ierr); 2478 if (!allcolumns[i]) { 2479 for (k=0; k<ncols; k++) { 2480 #if defined(PETSC_USE_CTABLE) 2481 ierr = PetscTableFind(cmap_i,cols[k]+1,&tcol);CHKERRQ(ierr); 2482 #else 2483 tcol = cmap_i[cols[k]]; 2484 #endif 2485 if (tcol) lens_i[j]++; 2486 } 2487 } else { /* allcolumns */ 2488 lens_i[j] = ncols; 2489 } 2490 ierr = MatRestoreRow_MPIAIJ(C,row,&ncols,&cols,0);CHKERRQ(ierr); 2491 } 2492 } 2493 } 2494 2495 /* Create row map: global row of C -> local row of submatrices */ 2496 #if defined(PETSC_USE_CTABLE) 2497 for (i=0; i<ismax; i++) { 2498 ierr = PetscTableCreate(nrow[i]+1,C->rmap->N+1,&rmap[i]);CHKERRQ(ierr); 2499 irow_i = irow[i]; 2500 jmax = nrow[i]; 2501 for (j=0; j<jmax; j++) { 2502 ierr = PetscTableAdd(rmap[i],irow_i[j]+1,j+1,INSERT_VALUES);CHKERRQ(ierr); 2503 } 2504 } 2505 #else 2506 if (ismax) { 2507 ierr = PetscMalloc1(ismax*C->rmap->N,&rmap[0]);CHKERRQ(ierr); 2508 ierr = PetscMemzero(rmap[0],ismax*C->rmap->N*sizeof(PetscInt));CHKERRQ(ierr); 2509 } 2510 for (i=1; i<ismax; i++) rmap[i] = rmap[i-1] + C->rmap->N; 2511 for (i=0; i<ismax; i++) { 2512 rmap_i = rmap[i]; 2513 irow_i = irow[i]; 2514 jmax = nrow[i]; 2515 for (j=0; j<jmax; j++) { 2516 rmap_i[irow_i[j]] = j; 2517 } 2518 } 2519 #endif 2520 2521 /* Update lens from offproc data */ 2522 { 2523 PetscInt *rbuf2_i,*rbuf3_i,*sbuf1_i; 2524 2525 ierr = MPI_Waitall(nrqs,r_waits3,r_status3);CHKERRQ(ierr); 2526 for (tmp2=0; tmp2<nrqs; tmp2++) { 2527 sbuf1_i = sbuf1[pa[tmp2]]; 2528 jmax = sbuf1_i[0]; 2529 ct1 = 2*jmax+1; 2530 ct2 = 0; 2531 rbuf2_i = rbuf2[tmp2]; 2532 rbuf3_i = rbuf3[tmp2]; 2533 for (j=1; j<=jmax; j++) { 2534 is_no = sbuf1_i[2*j-1]; 2535 max1 = sbuf1_i[2*j]; 2536 lens_i = lens[is_no]; 2537 if (!allcolumns[is_no]) cmap_i = cmap[is_no]; 2538 rmap_i = rmap[is_no]; 2539 for (k=0; k<max1; k++,ct1++) { 2540 #if defined(PETSC_USE_CTABLE) 2541 ierr = PetscTableFind(rmap_i,sbuf1_i[ct1]+1,&row);CHKERRQ(ierr); 2542 row--; 2543 if (row < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"row not found in table"); 2544 #else 2545 row = rmap_i[sbuf1_i[ct1]]; /* the val in the new matrix to be */ 2546 #endif 2547 max2 = rbuf2_i[ct1]; 2548 for (l=0; l<max2; l++,ct2++) { 2549 if (!allcolumns[is_no]) { 2550 #if defined(PETSC_USE_CTABLE) 2551 ierr = PetscTableFind(cmap_i,rbuf3_i[ct2]+1,&tcol);CHKERRQ(ierr); 2552 #else 2553 tcol = cmap_i[rbuf3_i[ct2]]; 2554 #endif 2555 if (tcol) lens_i[row]++; 2556 } else { /* allcolumns */ 2557 lens_i[row]++; /* lens_i[row] += max2 ? */ 2558 } 2559 } 2560 } 2561 } 2562 } 2563 } 2564 ierr = PetscFree(r_status3);CHKERRQ(ierr); 2565 ierr = PetscFree(r_waits3);CHKERRQ(ierr); 2566 if (nrqr) {ierr = MPI_Waitall(nrqr,s_waits3,s_status3);CHKERRQ(ierr);} 2567 ierr = PetscFree(s_status3);CHKERRQ(ierr); 2568 ierr = PetscFree(s_waits3);CHKERRQ(ierr); 2569 2570 /* Create the submatrices */ 2571 rmax = 0; 2572 for (i=0; i<ismax; i++) { 2573 PetscInt rbs,cbs; 2574 2575 ierr = ISGetBlockSize(isrow[i],&rbs);CHKERRQ(ierr); 2576 ierr = ISGetBlockSize(iscol[i],&cbs);CHKERRQ(ierr); 2577 2578 ierr = MatCreate(PETSC_COMM_SELF,submats+i);CHKERRQ(ierr); 2579 ierr = MatSetSizes(submats[i],nrow[i],ncol[i],PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 2580 2581 ierr = MatSetBlockSizes(submats[i],rbs,cbs);CHKERRQ(ierr); 2582 ierr = MatSetType(submats[i],((PetscObject)A)->type_name);CHKERRQ(ierr); 2583 ierr = MatSeqAIJSetPreallocation(submats[i],0,lens[i]);CHKERRQ(ierr); 2584 2585 /* create struct Mat_SubMat and attached it to submat */ 2586 ierr = PetscNew(&smat_i);CHKERRQ(ierr); 2587 subc = (Mat_SeqAIJ*)submats[i]->data; 2588 subc->submatis1 = smat_i; 2589 smats[i] = smat_i; 2590 2591 smat_i->destroy = submats[i]->ops->destroy; 2592 submats[i]->ops->destroy = MatDestroy_MPIAIJ_MatGetSubmatrices; 2593 submats[i]->factortype = C->factortype; 2594 2595 smat_i->id = i; 2596 smat_i->nrqs = nrqs; 2597 smat_i->nrqr = nrqr; 2598 smat_i->rbuf1 = rbuf1; 2599 smat_i->rbuf2 = rbuf2; 2600 smat_i->rbuf3 = rbuf3; 2601 smat_i->sbuf2 = sbuf2; 2602 smat_i->req_source2 = req_source2; 2603 2604 smat_i->sbuf1 = sbuf1; 2605 smat_i->ptr = ptr; 2606 smat_i->tmp = tmp; 2607 smat_i->ctr = ctr; 2608 2609 smat_i->pa = pa; 2610 smat_i->req_size = req_size; 2611 smat_i->req_source1 = req_source1; 2612 2613 smat_i->allcolumns = allcolumns[i]; 2614 smat_i->row2proc = row2proc[i]; 2615 smat_i->rmap = rmap[i]; 2616 smat_i->cmap = cmap[i]; 2617 #if defined(PETSC_USE_CTABLE) 2618 smat_i->rmap_loc = NULL;//rmap_loc; 2619 smat_i->cmap_loc = NULL;//cmap_loc; 2620 #endif 2621 2622 /* compute rmax */ 2623 lens_i = lens[i]; 2624 for (j=0; j<nrow[i]; j++) rmax = PetscMax(rmax,lens_i[j]); 2625 } 2626 //printf("[%d] rmax %d\n",rank,rmax); 2627 } /* endof scall == MAT_INITIAL_MATRIX */ 2628 2629 /* Post recv matrix values */ 2630 ierr = PetscObjectGetNewTag((PetscObject)C,&tag4);CHKERRQ(ierr); 2631 ierr = PetscMalloc1(nrqs+1,&rbuf4);CHKERRQ(ierr); 2632 ierr = PetscMalloc1(nrqs+1,&r_waits4);CHKERRQ(ierr); 2633 ierr = PetscMalloc1(nrqs+1,&r_status4);CHKERRQ(ierr); 2634 ierr = PetscMalloc1(nrqr+1,&s_status4);CHKERRQ(ierr); 2635 for (i=0; i<nrqs; ++i) { 2636 ierr = PetscMalloc1(rbuf2[i][0]+1,&rbuf4[i]);CHKERRQ(ierr); 2637 ierr = MPI_Irecv(rbuf4[i],rbuf2[i][0],MPIU_SCALAR,req_source2[i],tag4,comm,r_waits4+i);CHKERRQ(ierr); 2638 } 2639 2640 /* Allocate sending buffers for a->a, and send them off */ 2641 ierr = PetscMalloc1(nrqr+1,&sbuf_aa);CHKERRQ(ierr); 2642 for (i=0,j=0; i<nrqr; i++) j += req_size[i]; 2643 ierr = PetscMalloc1(j+1,&sbuf_aa[0]);CHKERRQ(ierr); 2644 for (i=1; i<nrqr; i++) sbuf_aa[i] = sbuf_aa[i-1] + req_size[i-1]; 2645 2646 ierr = PetscMalloc1(nrqr+1,&s_waits4);CHKERRQ(ierr); 2647 { 2648 PetscInt nzA,nzB,*a_i = a->i,*b_i = b->i, *cworkB,lwrite; 2649 PetscInt cstart = C->cmap->rstart,rstart = C->rmap->rstart,*bmap = c->garray; 2650 PetscInt cend = C->cmap->rend; 2651 PetscInt *b_j = b->j; 2652 PetscScalar *vworkA,*vworkB,*a_a = a->a,*b_a = b->a; 2653 2654 for (i=0; i<nrqr; i++) { 2655 rbuf1_i = rbuf1[i]; 2656 sbuf_aa_i = sbuf_aa[i]; 2657 ct1 = 2*rbuf1_i[0]+1; 2658 ct2 = 0; 2659 for (j=1,max1=rbuf1_i[0]; j<=max1; j++) { 2660 kmax = rbuf1_i[2*j]; 2661 for (k=0; k<kmax; k++,ct1++) { 2662 row = rbuf1_i[ct1] - rstart; 2663 nzA = a_i[row+1] - a_i[row]; nzB = b_i[row+1] - b_i[row]; 2664 ncols = nzA + nzB; 2665 cworkB = b_j + b_i[row]; 2666 vworkA = a_a + a_i[row]; 2667 vworkB = b_a + b_i[row]; 2668 2669 /* load the column values for this row into vals*/ 2670 vals = sbuf_aa_i+ct2; 2671 2672 lwrite = 0; 2673 for (l=0; l<nzB; l++) { 2674 if ((bmap[cworkB[l]]) < cstart) vals[lwrite++] = vworkB[l]; 2675 } 2676 for (l=0; l<nzA; l++) vals[lwrite++] = vworkA[l]; 2677 for (l=0; l<nzB; l++) { 2678 if ((bmap[cworkB[l]]) >= cend) vals[lwrite++] = vworkB[l]; 2679 } 2680 2681 ct2 += ncols; 2682 } 2683 } 2684 ierr = MPI_Isend(sbuf_aa_i,req_size[i],MPIU_SCALAR,req_source1[i],tag4,comm,s_waits4+i);CHKERRQ(ierr); 2685 } 2686 } 2687 2688 if (!ismax) { 2689 ierr = PetscFree(rbuf1[0]);CHKERRQ(ierr); 2690 ierr = PetscFree(rbuf1);CHKERRQ(ierr); 2691 } 2692 2693 /* Assemble the matrices */ 2694 /* First assemble the local rows */ 2695 { 2696 PetscInt ilen_row,*imat_ilen,*imat_j,*imat_i,old_row; 2697 PetscScalar *imat_a; 2698 2699 for (i=0; i<ismax; i++) { 2700 row2proc_i = row2proc[i]; 2701 subc = (Mat_SeqAIJ*)submats[i]->data; 2702 imat_ilen = subc->ilen; 2703 imat_j = subc->j; 2704 imat_i = subc->i; 2705 imat_a = subc->a; 2706 2707 if (!allcolumns[i]) cmap_i = cmap[i]; 2708 rmap_i = rmap[i]; 2709 irow_i = irow[i]; 2710 jmax = nrow[i]; 2711 for (j=0; j<jmax; j++) { 2712 row = irow_i[j]; 2713 proc = row2proc_i[j]; 2714 if (proc == rank) { 2715 old_row = row; 2716 #if defined(PETSC_USE_CTABLE) 2717 ierr = PetscTableFind(rmap_i,row+1,&row);CHKERRQ(ierr); 2718 row--; 2719 #else 2720 row = rmap_i[row]; 2721 #endif 2722 ilen_row = imat_ilen[row]; 2723 ierr = MatGetRow_MPIAIJ(C,old_row,&ncols,&cols,&vals);CHKERRQ(ierr); 2724 mat_i = imat_i[row]; 2725 mat_a = imat_a + mat_i; 2726 mat_j = imat_j + mat_i; 2727 if (!allcolumns[i]) { 2728 for (k=0; k<ncols; k++) { 2729 #if defined(PETSC_USE_CTABLE) 2730 ierr = PetscTableFind(cmap_i,cols[k]+1,&tcol);CHKERRQ(ierr); 2731 #else 2732 tcol = cmap_i[cols[k]]; 2733 #endif 2734 if (tcol) { 2735 *mat_j++ = tcol - 1; 2736 //if (rank==1) printf("%d - col %d\n",row,tcol - 1); 2737 *mat_a++ = vals[k]; 2738 ilen_row++; 2739 } 2740 } 2741 } else { /* allcolumns */ 2742 for (k=0; k<ncols; k++) { 2743 *mat_j++ = cols[k]; /* global col index! */ 2744 *mat_a++ = vals[k]; 2745 ilen_row++; 2746 } 2747 } 2748 ierr = MatRestoreRow_MPIAIJ(C,old_row,&ncols,&cols,&vals);CHKERRQ(ierr); 2749 2750 imat_ilen[row] = ilen_row; 2751 } 2752 } 2753 } 2754 } 2755 2756 /* Now assemble the off proc rows*/ 2757 { 2758 PetscInt *sbuf1_i,*rbuf2_i,*rbuf3_i,*imat_ilen,ilen; 2759 PetscInt *imat_j,*imat_i; 2760 PetscScalar *imat_a,*rbuf4_i; 2761 2762 ierr = MPI_Waitall(nrqs,r_waits4,r_status4);CHKERRQ(ierr); 2763 for (tmp2=0; tmp2<nrqs; tmp2++) { 2764 sbuf1_i = sbuf1[pa[tmp2]]; 2765 jmax = sbuf1_i[0]; 2766 ct1 = 2*jmax + 1; 2767 ct2 = 0; 2768 rbuf2_i = rbuf2[tmp2]; 2769 rbuf3_i = rbuf3[tmp2]; 2770 rbuf4_i = rbuf4[tmp2]; 2771 for (j=1; j<=jmax; j++) { 2772 is_no = sbuf1_i[2*j-1]; 2773 rmap_i = rmap[is_no]; 2774 if (!allcolumns[is_no]) cmap_i = cmap[is_no]; 2775 subc = (Mat_SeqAIJ*)submats[is_no]->data; 2776 imat_ilen = subc->ilen; 2777 imat_j = subc->j; 2778 imat_i = subc->i; 2779 imat_a = subc->a; 2780 max1 = sbuf1_i[2*j]; 2781 for (k=0; k<max1; k++,ct1++) { 2782 row = sbuf1_i[ct1]; 2783 #if defined(PETSC_USE_CTABLE) 2784 ierr = PetscTableFind(rmap_i,row+1,&row);CHKERRQ(ierr); 2785 row--; 2786 #else 2787 row = rmap_i[row]; 2788 #endif 2789 ilen = imat_ilen[row]; 2790 mat_i = imat_i[row]; 2791 mat_a = imat_a + mat_i; 2792 mat_j = imat_j + mat_i; 2793 max2 = rbuf2_i[ct1]; 2794 if (!allcolumns[is_no]) { 2795 for (l=0; l<max2; l++,ct2++) { 2796 2797 #if defined(PETSC_USE_CTABLE) 2798 ierr = PetscTableFind(cmap_i,rbuf3_i[ct2]+1,&tcol);CHKERRQ(ierr); 2799 #else 2800 tcol = cmap_i[rbuf3_i[ct2]]; 2801 #endif 2802 if (tcol) { 2803 *mat_j++ = tcol - 1; 2804 //if (rank==1) printf("%d - off-proc col %d\n",row,tcol - 1); 2805 *mat_a++ = rbuf4_i[ct2]; 2806 ilen++; 2807 } 2808 } 2809 } else { /* allcolumns */ 2810 for (l=0; l<max2; l++,ct2++) { 2811 *mat_j++ = rbuf3_i[ct2]; /* same global column index of C */ 2812 *mat_a++ = rbuf4_i[ct2]; 2813 ilen++; 2814 } 2815 } 2816 imat_ilen[row] = ilen; 2817 } 2818 } 2819 } 2820 } 2821 2822 /* sort the rows -- do we need this? */ 2823 { 2824 PetscInt *imat_ilen,*imat_j,*imat_i; 2825 PetscScalar *imat_a; 2826 2827 for (i=0; i<ismax; i++) { 2828 subc = (Mat_SeqAIJ*)submats[i]->data; 2829 imat_j = subc->j; 2830 imat_i = subc->i; 2831 imat_a = subc->a; 2832 imat_ilen = subc->ilen; 2833 2834 if (allcolumns[i]) continue; 2835 jmax = nrow[i]; 2836 for (j=0; j<jmax; j++) { 2837 PetscInt ilen; 2838 2839 mat_i = imat_i[j]; 2840 mat_a = imat_a + mat_i; 2841 mat_j = imat_j + mat_i; 2842 ilen = imat_ilen[j]; 2843 ierr = PetscSortIntWithScalarArray(ilen,mat_j,mat_a);CHKERRQ(ierr); 2844 } 2845 } 2846 } 2847 2848 ierr = PetscFree(r_status4);CHKERRQ(ierr); 2849 ierr = PetscFree(r_waits4);CHKERRQ(ierr); 2850 if (nrqr) {ierr = MPI_Waitall(nrqr,s_waits4,s_status4);CHKERRQ(ierr);} 2851 ierr = PetscFree(s_waits4);CHKERRQ(ierr); 2852 ierr = PetscFree(s_status4);CHKERRQ(ierr); 2853 2854 /* Restore the indices */ 2855 for (i=0; i<ismax; i++) { 2856 ierr = ISRestoreIndices(isrow[i],irow+i);CHKERRQ(ierr); 2857 if (!allcolumns[i]) { 2858 ierr = ISRestoreIndices(iscol[i],icol+i);CHKERRQ(ierr); 2859 } 2860 } 2861 2862 for (i=0; i<ismax; i++) { 2863 ierr = MatAssemblyBegin(submats[i],MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2864 ierr = MatAssemblyEnd(submats[i],MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2865 } 2866 2867 /* Destroy allocated memory */ 2868 if (scall == MAT_INITIAL_MATRIX) { 2869 ierr = PetscFree4(w1,w2,w3,w4);CHKERRQ(ierr); 2870 2871 if (ismax) {ierr = PetscFree(lens[0]);CHKERRQ(ierr);} 2872 ierr = PetscFree(lens);CHKERRQ(ierr); 2873 2874 ierr = PetscFree(sbuf_aj[0]);CHKERRQ(ierr); 2875 ierr = PetscFree(sbuf_aj);CHKERRQ(ierr); 2876 } 2877 2878 if (!ismax) { 2879 ierr = PetscFree(pa);CHKERRQ(ierr); 2880 2881 ierr = PetscFree4(sbuf1,ptr,tmp,ctr);CHKERRQ(ierr); 2882 for (i=0; i<nrqr; ++i) { 2883 ierr = PetscFree(sbuf2[i]);CHKERRQ(ierr); 2884 } 2885 for (i=0; i<nrqs; ++i) { 2886 ierr = PetscFree(rbuf3[i]);CHKERRQ(ierr); 2887 } 2888 2889 ierr = PetscFree3(sbuf2,req_size,req_source1);CHKERRQ(ierr); 2890 ierr = PetscFree3(req_source2,rbuf2,rbuf3);CHKERRQ(ierr); 2891 } 2892 2893 ierr = PetscFree(sbuf_aa[0]);CHKERRQ(ierr); 2894 ierr = PetscFree(sbuf_aa);CHKERRQ(ierr); 2895 2896 ierr = PetscFree4(irow,icol,nrow,ncol);CHKERRQ(ierr); 2897 2898 for (i=0; i<nrqs; ++i) { 2899 ierr = PetscFree(rbuf4[i]);CHKERRQ(ierr); 2900 } 2901 ierr = PetscFree(rbuf4);CHKERRQ(ierr); 2902 2903 ierr = PetscFree(rmap);CHKERRQ(ierr); 2904 ierr = PetscFree(cmap);CHKERRQ(ierr); 2905 2906 ierr = PetscFree(issorted);CHKERRQ(ierr); 2907 ierr = PetscFree(row2proc);CHKERRQ(ierr); 2908 ierr = PetscFree(smats);CHKERRQ(ierr); 2909 PetscFunctionReturn(0); 2910 } 2911 2912 /* 2913 Permute A & B into C's *local* index space using rowemb,dcolemb for A and rowemb,ocolemb for B. 2914 Embeddings are supposed to be injections and the above implies that the range of rowemb is a subset 2915 of [0,m), dcolemb is in [0,n) and ocolemb is in [N-n). 2916 If pattern == DIFFERENT_NONZERO_PATTERN, C is preallocated according to A&B. 2917 After that B's columns are mapped into C's global column space, so that C is in the "disassembled" 2918 state, and needs to be "assembled" later by compressing B's column space. 2919 2920 This function may be called in lieu of preallocation, so C should not be expected to be preallocated. 2921 Following this call, C->A & C->B have been created, even if empty. 2922 */ 2923 PetscErrorCode MatSetSeqMats_MPIAIJ(Mat C,IS rowemb,IS dcolemb,IS ocolemb,MatStructure pattern,Mat A,Mat B) 2924 { 2925 /* If making this function public, change the error returned in this function away from _PLIB. */ 2926 PetscErrorCode ierr; 2927 Mat_MPIAIJ *aij; 2928 Mat_SeqAIJ *Baij; 2929 PetscBool seqaij,Bdisassembled; 2930 PetscInt m,n,*nz,i,j,ngcol,col,rstart,rend,shift,count; 2931 PetscScalar v; 2932 const PetscInt *rowindices,*colindices; 2933 2934 PetscFunctionBegin; 2935 /* Check to make sure the component matrices (and embeddings) are compatible with C. */ 2936 if (A) { 2937 ierr = PetscObjectTypeCompare((PetscObject)A,MATSEQAIJ,&seqaij);CHKERRQ(ierr); 2938 if (!seqaij) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Diagonal matrix is of wrong type"); 2939 if (rowemb) { 2940 ierr = ISGetLocalSize(rowemb,&m);CHKERRQ(ierr); 2941 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); 2942 } else { 2943 if (C->rmap->n != A->rmap->n) { 2944 SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Diag seq matrix is row-incompatible with the MPIAIJ matrix"); 2945 } 2946 } 2947 if (dcolemb) { 2948 ierr = ISGetLocalSize(dcolemb,&n);CHKERRQ(ierr); 2949 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); 2950 } else { 2951 if (C->cmap->n != A->cmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Diag seq matrix is col-incompatible with the MPIAIJ matrix"); 2952 } 2953 } 2954 if (B) { 2955 ierr = PetscObjectTypeCompare((PetscObject)B,MATSEQAIJ,&seqaij);CHKERRQ(ierr); 2956 if (!seqaij) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Off-diagonal matrix is of wrong type"); 2957 if (rowemb) { 2958 ierr = ISGetLocalSize(rowemb,&m);CHKERRQ(ierr); 2959 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); 2960 } else { 2961 if (C->rmap->n != B->rmap->n) { 2962 SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Off-diag seq matrix is row-incompatible with the MPIAIJ matrix"); 2963 } 2964 } 2965 if (ocolemb) { 2966 ierr = ISGetLocalSize(ocolemb,&n);CHKERRQ(ierr); 2967 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); 2968 } else { 2969 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"); 2970 } 2971 } 2972 2973 aij = (Mat_MPIAIJ*)(C->data); 2974 if (!aij->A) { 2975 /* Mimic parts of MatMPIAIJSetPreallocation() */ 2976 ierr = MatCreate(PETSC_COMM_SELF,&aij->A);CHKERRQ(ierr); 2977 ierr = MatSetSizes(aij->A,C->rmap->n,C->cmap->n,C->rmap->n,C->cmap->n);CHKERRQ(ierr); 2978 ierr = MatSetBlockSizesFromMats(aij->A,C,C);CHKERRQ(ierr); 2979 ierr = MatSetType(aij->A,MATSEQAIJ);CHKERRQ(ierr); 2980 ierr = PetscLogObjectParent((PetscObject)C,(PetscObject)aij->A);CHKERRQ(ierr); 2981 } 2982 if (A) { 2983 ierr = MatSetSeqMat_SeqAIJ(aij->A,rowemb,dcolemb,pattern,A);CHKERRQ(ierr); 2984 } else { 2985 ierr = MatSetUp(aij->A);CHKERRQ(ierr); 2986 } 2987 if (B) { /* Destroy the old matrix or the column map, depending on the sparsity pattern. */ 2988 /* 2989 If pattern == DIFFERENT_NONZERO_PATTERN, we reallocate B and 2990 need to "disassemble" B -- convert it to using C's global indices. 2991 To insert the values we take the safer, albeit more expensive, route of MatSetValues(). 2992 2993 If pattern == SUBSET_NONZERO_PATTERN, we do not "disassemble" B and do not reallocate; 2994 we MatZeroValues(B) first, so there may be a bunch of zeros that, perhaps, could be compacted out. 2995 2996 TODO: Put B's values into aij->B's aij structure in place using the embedding ISs? 2997 At least avoid calling MatSetValues() and the implied searches? 2998 */ 2999 3000 if (B && pattern == DIFFERENT_NONZERO_PATTERN) { 3001 #if defined(PETSC_USE_CTABLE) 3002 ierr = PetscTableDestroy(&aij->colmap);CHKERRQ(ierr); 3003 #else 3004 ierr = PetscFree(aij->colmap);CHKERRQ(ierr); 3005 /* A bit of a HACK: ideally we should deal with case aij->B all in one code block below. */ 3006 if (aij->B) { 3007 ierr = PetscLogObjectMemory((PetscObject)C,-aij->B->cmap->n*sizeof(PetscInt));CHKERRQ(ierr); 3008 } 3009 #endif 3010 ngcol = 0; 3011 if (aij->lvec) { 3012 ierr = VecGetSize(aij->lvec,&ngcol);CHKERRQ(ierr); 3013 } 3014 if (aij->garray) { 3015 ierr = PetscFree(aij->garray);CHKERRQ(ierr); 3016 ierr = PetscLogObjectMemory((PetscObject)C,-ngcol*sizeof(PetscInt));CHKERRQ(ierr); 3017 } 3018 ierr = VecDestroy(&aij->lvec);CHKERRQ(ierr); 3019 ierr = VecScatterDestroy(&aij->Mvctx);CHKERRQ(ierr); 3020 } 3021 if (aij->B && B && pattern == DIFFERENT_NONZERO_PATTERN) { 3022 ierr = MatDestroy(&aij->B);CHKERRQ(ierr); 3023 } 3024 if (aij->B && B && pattern == SUBSET_NONZERO_PATTERN) { 3025 ierr = MatZeroEntries(aij->B);CHKERRQ(ierr); 3026 } 3027 } 3028 Bdisassembled = PETSC_FALSE; 3029 if (!aij->B) { 3030 ierr = MatCreate(PETSC_COMM_SELF,&aij->B);CHKERRQ(ierr); 3031 ierr = PetscLogObjectParent((PetscObject)C,(PetscObject)aij->B);CHKERRQ(ierr); 3032 ierr = MatSetSizes(aij->B,C->rmap->n,C->cmap->N,C->rmap->n,C->cmap->N);CHKERRQ(ierr); 3033 ierr = MatSetBlockSizesFromMats(aij->B,B,B);CHKERRQ(ierr); 3034 ierr = MatSetType(aij->B,MATSEQAIJ);CHKERRQ(ierr); 3035 Bdisassembled = PETSC_TRUE; 3036 } 3037 if (B) { 3038 Baij = (Mat_SeqAIJ*)(B->data); 3039 if (pattern == DIFFERENT_NONZERO_PATTERN) { 3040 ierr = PetscMalloc1(B->rmap->n,&nz);CHKERRQ(ierr); 3041 for (i=0; i<B->rmap->n; i++) { 3042 nz[i] = Baij->i[i+1] - Baij->i[i]; 3043 } 3044 ierr = MatSeqAIJSetPreallocation(aij->B,0,nz);CHKERRQ(ierr); 3045 ierr = PetscFree(nz);CHKERRQ(ierr); 3046 } 3047 3048 ierr = PetscLayoutGetRange(C->rmap,&rstart,&rend);CHKERRQ(ierr); 3049 shift = rend-rstart; 3050 count = 0; 3051 rowindices = NULL; 3052 colindices = NULL; 3053 if (rowemb) { 3054 ierr = ISGetIndices(rowemb,&rowindices);CHKERRQ(ierr); 3055 } 3056 if (ocolemb) { 3057 ierr = ISGetIndices(ocolemb,&colindices);CHKERRQ(ierr); 3058 } 3059 for (i=0; i<B->rmap->n; i++) { 3060 PetscInt row; 3061 row = i; 3062 if (rowindices) row = rowindices[i]; 3063 for (j=Baij->i[i]; j<Baij->i[i+1]; j++) { 3064 col = Baij->j[count]; 3065 if (colindices) col = colindices[col]; 3066 if (Bdisassembled && col>=rstart) col += shift; 3067 v = Baij->a[count]; 3068 ierr = MatSetValues(aij->B,1,&row,1,&col,&v,INSERT_VALUES);CHKERRQ(ierr); 3069 ++count; 3070 } 3071 } 3072 /* No assembly for aij->B is necessary. */ 3073 /* FIXME: set aij->B's nonzerostate correctly. */ 3074 } else { 3075 ierr = MatSetUp(aij->B);CHKERRQ(ierr); 3076 } 3077 C->preallocated = PETSC_TRUE; 3078 C->was_assembled = PETSC_FALSE; 3079 C->assembled = PETSC_FALSE; 3080 /* 3081 C will need to be assembled so that aij->B can be compressed into local form in MatSetUpMultiply_MPIAIJ(). 3082 Furthermore, its nonzerostate will need to be based on that of aij->A's and aij->B's. 3083 */ 3084 PetscFunctionReturn(0); 3085 } 3086 3087 /* 3088 B uses local indices with column indices ranging between 0 and N-n; they must be interpreted using garray. 3089 */ 3090 PetscErrorCode MatGetSeqMats_MPIAIJ(Mat C,Mat *A,Mat *B) 3091 { 3092 Mat_MPIAIJ *aij = (Mat_MPIAIJ*) (C->data); 3093 3094 PetscFunctionBegin; 3095 PetscValidPointer(A,2); 3096 PetscValidPointer(B,3); 3097 /* FIXME: make sure C is assembled */ 3098 *A = aij->A; 3099 *B = aij->B; 3100 /* Note that we don't incref *A and *B, so be careful! */ 3101 PetscFunctionReturn(0); 3102 } 3103 3104 /* 3105 Extract MPI submatrices encoded by pairs of IS that may live on subcomms of C. 3106 NOT SCALABLE due to the use of ISGetNonlocalIS() (see below). 3107 */ 3108 PetscErrorCode MatGetSubMatricesMPI_MPIXAIJ(Mat C,PetscInt ismax,const IS isrow[],const IS iscol[],MatReuse scall,Mat *submat[], 3109 PetscErrorCode(*getsubmats_seq)(Mat,PetscInt,const IS[],const IS[],MatReuse,Mat**), 3110 PetscErrorCode(*getlocalmats)(Mat,Mat*,Mat*), 3111 PetscErrorCode(*setseqmat)(Mat,IS,IS,MatStructure,Mat), 3112 PetscErrorCode(*setseqmats)(Mat,IS,IS,IS,MatStructure,Mat,Mat)) 3113 { 3114 PetscErrorCode ierr; 3115 PetscMPIInt isize,flag; 3116 PetscInt i,ii,cismax,ispar; 3117 Mat *A,*B; 3118 IS *isrow_p,*iscol_p,*cisrow,*ciscol,*ciscol_p; 3119 3120 PetscFunctionBegin; 3121 if (!ismax) PetscFunctionReturn(0); 3122 3123 for (i = 0, cismax = 0; i < ismax; ++i) { 3124 PetscMPIInt isize; 3125 ierr = MPI_Comm_compare(((PetscObject)isrow[i])->comm,((PetscObject)iscol[i])->comm,&flag);CHKERRQ(ierr); 3126 if (flag != MPI_IDENT) SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Row and column index sets must have the same communicator"); 3127 ierr = MPI_Comm_size(((PetscObject)isrow[i])->comm, &isize);CHKERRQ(ierr); 3128 if (isize > 1) ++cismax; 3129 } 3130 3131 /* 3132 If cismax is zero on all C's ranks, then and only then can we use purely sequential matrix extraction. 3133 ispar counts the number of parallel ISs across C's comm. 3134 */ 3135 ierr = MPIU_Allreduce(&cismax,&ispar,1,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)C));CHKERRQ(ierr); 3136 if (!ispar) { /* Sequential ISs only across C's comm, so can call the sequential matrix extraction subroutine. */ 3137 ierr = (*getsubmats_seq)(C,ismax,isrow,iscol,scall,submat);CHKERRQ(ierr); 3138 PetscFunctionReturn(0); 3139 } 3140 3141 /* if (ispar) */ 3142 /* 3143 Construct the "complements" -- the off-processor indices -- of the iscol ISs for parallel ISs only. 3144 These are used to extract the off-diag portion of the resulting parallel matrix. 3145 The row IS for the off-diag portion is the same as for the diag portion, 3146 so we merely alias (without increfing) the row IS, while skipping those that are sequential. 3147 */ 3148 ierr = PetscMalloc2(cismax,&cisrow,cismax,&ciscol);CHKERRQ(ierr); 3149 ierr = PetscMalloc1(cismax,&ciscol_p);CHKERRQ(ierr); 3150 for (i = 0, ii = 0; i < ismax; ++i) { 3151 ierr = MPI_Comm_size(((PetscObject)isrow[i])->comm,&isize);CHKERRQ(ierr); 3152 if (isize > 1) { 3153 /* 3154 TODO: This is the part that's ***NOT SCALABLE***. 3155 To fix this we need to extract just the indices of C's nonzero columns 3156 that lie on the intersection of isrow[i] and ciscol[ii] -- the nonlocal 3157 part of iscol[i] -- without actually computing ciscol[ii]. This also has 3158 to be done without serializing on the IS list, so, most likely, it is best 3159 done by rewriting MatGetSubMatrices_MPIAIJ() directly. 3160 */ 3161 ierr = ISGetNonlocalIS(iscol[i],&(ciscol[ii]));CHKERRQ(ierr); 3162 /* Now we have to 3163 (a) make sure ciscol[ii] is sorted, since, even if the off-proc indices 3164 were sorted on each rank, concatenated they might no longer be sorted; 3165 (b) Use ISSortPermutation() to construct ciscol_p, the mapping from the 3166 indices in the nondecreasing order to the original index positions. 3167 If ciscol[ii] is strictly increasing, the permutation IS is NULL. 3168 */ 3169 ierr = ISSortPermutation(ciscol[ii],PETSC_FALSE,ciscol_p+ii);CHKERRQ(ierr); 3170 ierr = ISSort(ciscol[ii]);CHKERRQ(ierr); 3171 ++ii; 3172 } 3173 } 3174 ierr = PetscMalloc2(ismax,&isrow_p,ismax,&iscol_p);CHKERRQ(ierr); 3175 for (i = 0, ii = 0; i < ismax; ++i) { 3176 PetscInt j,issize; 3177 const PetscInt *indices; 3178 3179 /* 3180 Permute the indices into a nondecreasing order. Reject row and col indices with duplicates. 3181 */ 3182 ierr = ISSortPermutation(isrow[i],PETSC_FALSE,isrow_p+i);CHKERRQ(ierr); 3183 ierr = ISSort(isrow[i]);CHKERRQ(ierr); 3184 ierr = ISGetLocalSize(isrow[i],&issize);CHKERRQ(ierr); 3185 ierr = ISGetIndices(isrow[i],&indices);CHKERRQ(ierr); 3186 for (j = 1; j < issize; ++j) { 3187 if (indices[j] == indices[j-1]) { 3188 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]); 3189 } 3190 } 3191 ierr = ISRestoreIndices(isrow[i],&indices);CHKERRQ(ierr); 3192 3193 3194 ierr = ISSortPermutation(iscol[i],PETSC_FALSE,iscol_p+i);CHKERRQ(ierr); 3195 ierr = ISSort(iscol[i]);CHKERRQ(ierr); 3196 ierr = ISGetLocalSize(iscol[i],&issize);CHKERRQ(ierr); 3197 ierr = ISGetIndices(iscol[i],&indices);CHKERRQ(ierr); 3198 for (j = 1; j < issize; ++j) { 3199 if (indices[j-1] == indices[j]) { 3200 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]); 3201 } 3202 } 3203 ierr = ISRestoreIndices(iscol[i],&indices);CHKERRQ(ierr); 3204 ierr = MPI_Comm_size(((PetscObject)isrow[i])->comm,&isize);CHKERRQ(ierr); 3205 if (isize > 1) { 3206 cisrow[ii] = isrow[i]; 3207 ++ii; 3208 } 3209 } 3210 /* 3211 Allocate the necessary arrays to hold the resulting parallel matrices as well as the intermediate 3212 array of sequential matrices underlying the resulting parallel matrices. 3213 Which arrays to allocate is based on the value of MatReuse scall and whether ISs are sorted and/or 3214 contain duplicates. 3215 3216 There are as many diag matrices as there are original index sets. There are only as many parallel 3217 and off-diag matrices, as there are parallel (comm size > 1) index sets. 3218 3219 ARRAYS that can hold Seq matrices get allocated in any event -- either here or by getsubmats_seq(): 3220 - If the array of MPI matrices already exists and is being reused, we need to allocate the array 3221 and extract the underlying seq matrices into it to serve as placeholders, into which getsubmats_seq 3222 will deposite the extracted diag and off-diag parts. Thus, we allocate the A&B arrays and fill them 3223 with A[i] and B[ii] extracted from the corresponding MPI submat. 3224 - However, if the rows, A's column indices or B's column indices are not sorted, the extracted A[i] & B[ii] 3225 will have a different order from what getsubmats_seq expects. To handle this case -- indicated 3226 by a nonzero isrow_p[i], iscol_p[i], or ciscol_p[ii] -- we duplicate A[i] --> AA[i], B[ii] --> BB[ii] 3227 (retrieve composed AA[i] or BB[ii]) and reuse them here. AA[i] and BB[ii] are then used to permute its 3228 values into A[i] and B[ii] sitting inside the corresponding submat. 3229 - If no reuse is taking place then getsubmats_seq will allocate the A&B arrays and create the corresponding 3230 A[i], B[ii], AA[i] or BB[ii] matrices. 3231 */ 3232 /* Parallel matrix array is allocated here only if no reuse is taking place. If reused, it is passed in by the caller. */ 3233 if (scall == MAT_INITIAL_MATRIX) { 3234 ierr = PetscMalloc1(ismax,submat);CHKERRQ(ierr); 3235 } 3236 3237 /* Now obtain the sequential A and B submatrices separately. */ 3238 /* scall=MAT_REUSE_MATRIX is not handled yet, because getsubmats_seq() requires reuse of A and B */ 3239 ierr = (*getsubmats_seq)(C,ismax,isrow,iscol,MAT_INITIAL_MATRIX,&A);CHKERRQ(ierr); 3240 ierr = (*getsubmats_seq)(C,cismax,cisrow,ciscol,MAT_INITIAL_MATRIX,&B);CHKERRQ(ierr); 3241 3242 /* 3243 If scall == MAT_REUSE_MATRIX AND the permutations are NULL, we are done, since the sequential 3244 matrices A & B have been extracted directly into the parallel matrices containing them, or 3245 simply into the sequential matrix identical with the corresponding A (if isize == 1). 3246 Note that in that case colmap doesn't need to be rebuilt, since the matrices are expected 3247 to have the same sparsity pattern. 3248 Otherwise, A and/or B have to be properly embedded into C's index spaces and the correct colmap 3249 must be constructed for C. This is done by setseqmat(s). 3250 */ 3251 for (i = 0, ii = 0; i < ismax; ++i) { 3252 /* 3253 TODO: cache ciscol, permutation ISs and maybe cisrow? What about isrow & iscol? 3254 That way we can avoid sorting and computing permutations when reusing. 3255 To this end: 3256 - remove the old cache, if it exists, when extracting submatrices with MAT_INITIAL_MATRIX 3257 - if caching arrays to hold the ISs, make and compose a container for them so that it can 3258 be destroyed upon destruction of C (use PetscContainerUserDestroy() to clear out the contents). 3259 */ 3260 MatStructure pattern; 3261 pattern = DIFFERENT_NONZERO_PATTERN; 3262 3263 ierr = MPI_Comm_size(((PetscObject)isrow[i])->comm,&isize);CHKERRQ(ierr); 3264 /* Construct submat[i] from the Seq pieces A (and B, if necessary). */ 3265 if (isize > 1) { 3266 if (scall == MAT_INITIAL_MATRIX) { 3267 ierr = MatCreate(((PetscObject)isrow[i])->comm,(*submat)+i);CHKERRQ(ierr); 3268 ierr = MatSetSizes((*submat)[i],A[i]->rmap->n,A[i]->cmap->n,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 3269 ierr = MatSetType((*submat)[i],MATMPIAIJ);CHKERRQ(ierr); 3270 ierr = PetscLayoutSetUp((*submat)[i]->rmap);CHKERRQ(ierr); 3271 ierr = PetscLayoutSetUp((*submat)[i]->cmap);CHKERRQ(ierr); 3272 } 3273 /* 3274 For each parallel isrow[i], insert the extracted sequential matrices into the parallel matrix. 3275 */ 3276 { 3277 Mat AA,BB; 3278 AA = A[i]; 3279 BB = B[ii]; 3280 if (AA || BB) { 3281 ierr = setseqmats((*submat)[i],isrow_p[i],iscol_p[i],ciscol_p[ii],pattern,AA,BB);CHKERRQ(ierr); 3282 ierr = MatAssemblyBegin((*submat)[i],MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3283 ierr = MatAssemblyEnd((*submat)[i],MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3284 } 3285 3286 ierr = MatDestroy(&AA);CHKERRQ(ierr); 3287 ierr = MatDestroy(&BB);CHKERRQ(ierr); 3288 } 3289 ierr = ISDestroy(ciscol+ii);CHKERRQ(ierr); 3290 ierr = ISDestroy(ciscol_p+ii);CHKERRQ(ierr); 3291 ++ii; 3292 } else { /* if (isize == 1) */ 3293 if (scall == MAT_REUSE_MATRIX) { 3294 ierr = MatDestroy(&(*submat)[i]);CHKERRQ(ierr); 3295 } 3296 if (isrow_p[i] || iscol_p[i]) { 3297 ierr = MatDuplicate(A[i],MAT_DO_NOT_COPY_VALUES,(*submat)+i);CHKERRQ(ierr); 3298 ierr = setseqmat((*submat)[i],isrow_p[i],iscol_p[i],pattern,A[i]);CHKERRQ(ierr); 3299 /* Otherwise A is extracted straight into (*submats)[i]. */ 3300 /* TODO: Compose A[i] on (*submat([i] for future use, if ((isrow_p[i] || iscol_p[i]) && MAT_INITIAL_MATRIX). */ 3301 ierr = MatDestroy(A+i);CHKERRQ(ierr); 3302 } else (*submat)[i] = A[i]; 3303 } 3304 ierr = ISDestroy(&isrow_p[i]);CHKERRQ(ierr); 3305 ierr = ISDestroy(&iscol_p[i]);CHKERRQ(ierr); 3306 } 3307 ierr = PetscFree2(cisrow,ciscol);CHKERRQ(ierr); 3308 ierr = PetscFree2(isrow_p,iscol_p);CHKERRQ(ierr); 3309 ierr = PetscFree(ciscol_p);CHKERRQ(ierr); 3310 ierr = PetscFree(A);CHKERRQ(ierr); 3311 ierr = PetscFree(B);CHKERRQ(ierr); 3312 PetscFunctionReturn(0); 3313 } 3314 3315 PetscErrorCode MatGetSubMatricesMPI_MPIAIJ(Mat C,PetscInt ismax,const IS isrow[],const IS iscol[],MatReuse scall,Mat *submat[]) 3316 { 3317 PetscErrorCode ierr; 3318 3319 PetscFunctionBegin; 3320 ierr = MatGetSubMatricesMPI_MPIXAIJ(C,ismax,isrow,iscol,scall,submat,MatGetSubMatrices_MPIAIJ,MatGetSeqMats_MPIAIJ,MatSetSeqMat_SeqAIJ,MatSetSeqMats_MPIAIJ);CHKERRQ(ierr); 3321 PetscFunctionReturn(0); 3322 } 3323