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 ierr = submatj->destroy(C);CHKERRQ(ierr); 1232 1233 ierr = PetscFree4(submatj->sbuf1,submatj->ptr,submatj->tmp,submatj->ctr);CHKERRQ(ierr); 1234 1235 for (i=0; i<submatj->nrqr; ++i) { 1236 ierr = PetscFree(submatj->sbuf2[i]);CHKERRQ(ierr); 1237 } 1238 ierr = PetscFree3(submatj->sbuf2,submatj->req_size,submatj->req_source1);CHKERRQ(ierr); 1239 1240 if (submatj->rbuf1) { 1241 ierr = PetscFree(submatj->rbuf1[0]);CHKERRQ(ierr); 1242 ierr = PetscFree(submatj->rbuf1);CHKERRQ(ierr); 1243 } 1244 1245 for (i=0; i<submatj->nrqs; ++i) { 1246 ierr = PetscFree(submatj->rbuf3[i]);CHKERRQ(ierr); 1247 } 1248 ierr = PetscFree3(submatj->req_source2,submatj->rbuf2,submatj->rbuf3);CHKERRQ(ierr); 1249 ierr = PetscFree(submatj->pa);CHKERRQ(ierr); 1250 1251 ierr = PetscFree(submatj->row2proc);CHKERRQ(ierr); 1252 1253 #if defined(PETSC_USE_CTABLE) 1254 ierr = PetscTableDestroy((PetscTable*)&submatj->rmap);CHKERRQ(ierr); 1255 if (submatj->cmap_loc) {ierr = PetscFree(submatj->cmap_loc);CHKERRQ(ierr);} 1256 ierr = PetscFree(submatj->rmap_loc);CHKERRQ(ierr); 1257 #else 1258 ierr = PetscFree(submatj->rmap);CHKERRQ(ierr); 1259 #endif 1260 1261 if (!submatj->allcolumns) { 1262 #if defined(PETSC_USE_CTABLE) 1263 ierr = PetscTableDestroy((PetscTable*)&submatj->cmap);CHKERRQ(ierr); 1264 #else 1265 ierr = PetscFree(submatj->cmap);CHKERRQ(ierr); 1266 #endif 1267 } 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->nrqs = nrqs; 1683 smatis1->nrqr = nrqr; 1684 smatis1->rbuf1 = rbuf1; 1685 smatis1->rbuf2 = rbuf2; 1686 smatis1->rbuf3 = rbuf3; 1687 smatis1->sbuf2 = sbuf2; 1688 smatis1->req_source2 = req_source2; 1689 1690 smatis1->sbuf1 = sbuf1; 1691 smatis1->ptr = ptr; 1692 smatis1->tmp = tmp; 1693 smatis1->ctr = ctr; 1694 1695 smatis1->pa = pa; 1696 smatis1->req_size = req_size; 1697 smatis1->req_source1 = req_source1; 1698 1699 smatis1->allcolumns = allcolumns; 1700 smatis1->row2proc = row2proc; 1701 smatis1->rmap = rmap; 1702 smatis1->cmap = cmap; 1703 #if defined(PETSC_USE_CTABLE) 1704 smatis1->rmap_loc = rmap_loc; 1705 smatis1->cmap_loc = cmap_loc; 1706 #endif 1707 1708 smatis1->destroy = submat->ops->destroy; 1709 submat->ops->destroy = MatDestroy_MPIAIJ_MatGetSubmatrices; 1710 submat->factortype = C->factortype; 1711 1712 /* compute rmax */ 1713 rmax = 0; 1714 for (i=0; i<nrow; i++) rmax = PetscMax(rmax,lens[i]); 1715 1716 } else { /* scall == MAT_REUSE_MATRIX */ 1717 submat = submats[0]; 1718 if (submat->rmap->n != nrow || submat->cmap->n != ncol) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Cannot reuse matrix. wrong size"); 1719 1720 subc = (Mat_SeqAIJ*)submat->data; 1721 rmax = subc->rmax; 1722 smatis1 = subc->submatis1; 1723 nrqs = smatis1->nrqs; 1724 nrqr = smatis1->nrqr; 1725 rbuf1 = smatis1->rbuf1; 1726 rbuf2 = smatis1->rbuf2; 1727 rbuf3 = smatis1->rbuf3; 1728 req_source2 = smatis1->req_source2; 1729 1730 sbuf1 = smatis1->sbuf1; 1731 sbuf2 = smatis1->sbuf2; 1732 ptr = smatis1->ptr; 1733 tmp = smatis1->tmp; 1734 ctr = smatis1->ctr; 1735 1736 pa = smatis1->pa; 1737 req_size = smatis1->req_size; 1738 req_source1 = smatis1->req_source1; 1739 1740 allcolumns = smatis1->allcolumns; 1741 row2proc = smatis1->row2proc; 1742 rmap = smatis1->rmap; 1743 cmap = smatis1->cmap; 1744 #if defined(PETSC_USE_CTABLE) 1745 rmap_loc = smatis1->rmap_loc; 1746 cmap_loc = smatis1->cmap_loc; 1747 #endif 1748 } 1749 1750 /* Post recv matrix values */ 1751 ierr = PetscMalloc3(nrqs+1,&rbuf4, rmax,&subcols, rmax,&subvals);CHKERRQ(ierr); 1752 ierr = PetscMalloc4(nrqs+1,&r_waits4,nrqr+1,&s_waits4,nrqs+1,&r_status4,nrqr+1,&s_status4);CHKERRQ(ierr); 1753 ierr = PetscObjectGetNewTag((PetscObject)C,&tag4);CHKERRQ(ierr); 1754 for (i=0; i<nrqs; ++i) { 1755 ierr = PetscMalloc1(rbuf2[i][0]+1,&rbuf4[i]);CHKERRQ(ierr); 1756 ierr = MPI_Irecv(rbuf4[i],rbuf2[i][0],MPIU_SCALAR,req_source2[i],tag4,comm,r_waits4+i);CHKERRQ(ierr); 1757 } 1758 1759 /* Allocate sending buffers for a->a, and send them off */ 1760 ierr = PetscMalloc1(nrqr+1,&sbuf_aa);CHKERRQ(ierr); 1761 for (i=0,j=0; i<nrqr; i++) j += req_size[i]; 1762 ierr = PetscMalloc1(j+1,&sbuf_aa[0]);CHKERRQ(ierr); 1763 for (i=1; i<nrqr; i++) sbuf_aa[i] = sbuf_aa[i-1] + req_size[i-1]; 1764 1765 for (i=0; i<nrqr; i++) { 1766 rbuf1_i = rbuf1[i]; 1767 sbuf_aa_i = sbuf_aa[i]; 1768 ct1 = 2*rbuf1_i[0]+1; 1769 ct2 = 0; 1770 /* max1=rbuf1_i[0]; if (max1 != 1) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"max1 !=1"); */ 1771 1772 kmax = rbuf1_i[2]; 1773 for (k=0; k<kmax; k++,ct1++) { 1774 row = rbuf1_i[ct1] - rstart; 1775 nzA = ai[row+1] - ai[row]; 1776 nzB = bi[row+1] - bi[row]; 1777 ncols = nzA + nzB; 1778 cworkB = bj + bi[row]; 1779 vworkA = a_a + ai[row]; 1780 vworkB = b_a + bi[row]; 1781 1782 /* load the column values for this row into vals*/ 1783 vals = sbuf_aa_i + ct2; 1784 1785 lwrite = 0; 1786 for (l=0; l<nzB; l++) { 1787 if ((bmap[cworkB[l]]) < cstart) vals[lwrite++] = vworkB[l]; 1788 } 1789 for (l=0; l<nzA; l++) vals[lwrite++] = vworkA[l]; 1790 for (l=0; l<nzB; l++) { 1791 if ((bmap[cworkB[l]]) >= cend) vals[lwrite++] = vworkB[l]; 1792 } 1793 1794 ct2 += ncols; 1795 } 1796 ierr = MPI_Isend(sbuf_aa_i,req_size[i],MPIU_SCALAR,req_source1[i],tag4,comm,s_waits4+i);CHKERRQ(ierr); 1797 } 1798 1799 /* Assemble submat */ 1800 /* First assemble the local rows */ 1801 for (j=0; j<nrow; j++) { 1802 row = irow[j]; 1803 proc = row2proc[j]; 1804 if (proc == rank) { 1805 Crow = row - rstart; /* local row index of C */ 1806 #if defined(PETSC_USE_CTABLE) 1807 row = rmap_loc[Crow]; /* row index of submat */ 1808 #else 1809 row = rmap[row]; 1810 #endif 1811 1812 if (allcolumns) { 1813 /* diagonal part A = c->A */ 1814 ncols = ai[Crow+1] - ai[Crow]; 1815 cols = aj + ai[Crow]; 1816 vals = a->a + ai[Crow]; 1817 i = 0; 1818 for (k=0; k<ncols; k++) { 1819 subcols[i] = cols[k] + cstart; 1820 subvals[i++] = vals[k]; 1821 } 1822 1823 /* off-diagonal part B = c->B */ 1824 ncols = bi[Crow+1] - bi[Crow]; 1825 cols = bj + bi[Crow]; 1826 vals = b->a + bi[Crow]; 1827 for (k=0; k<ncols; k++) { 1828 subcols[i] = bmap[cols[k]]; 1829 subvals[i++] = vals[k]; 1830 } 1831 1832 ierr = MatSetValues_SeqAIJ(submat,1,&row,i,subcols,subvals,INSERT_VALUES);CHKERRQ(ierr); 1833 1834 } else { /* !allcolumns */ 1835 #if defined(PETSC_USE_CTABLE) 1836 /* diagonal part A = c->A */ 1837 ncols = ai[Crow+1] - ai[Crow]; 1838 cols = aj + ai[Crow]; 1839 vals = a->a + ai[Crow]; 1840 i = 0; 1841 for (k=0; k<ncols; k++) { 1842 tcol = cmap_loc[cols[k]]; 1843 if (tcol) { 1844 subcols[i] = --tcol; 1845 subvals[i++] = vals[k]; 1846 } 1847 } 1848 1849 /* off-diagonal part B = c->B */ 1850 ncols = bi[Crow+1] - bi[Crow]; 1851 cols = bj + bi[Crow]; 1852 vals = b->a + bi[Crow]; 1853 for (k=0; k<ncols; k++) { 1854 ierr = PetscTableFind(cmap,bmap[cols[k]]+1,&tcol);CHKERRQ(ierr); 1855 if (tcol) { 1856 subcols[i] = --tcol; 1857 subvals[i++] = vals[k]; 1858 } 1859 } 1860 #else 1861 /* diagonal part A = c->A */ 1862 ncols = ai[Crow+1] - ai[Crow]; 1863 cols = aj + ai[Crow]; 1864 vals = a->a + ai[Crow]; 1865 i = 0; 1866 for (k=0; k<ncols; k++) { 1867 tcol = cmap[cols[k]+cstart]; 1868 if (tcol) { 1869 subcols[i] = --tcol; 1870 subvals[i++] = vals[k]; 1871 } 1872 } 1873 1874 /* off-diagonal part B = c->B */ 1875 ncols = bi[Crow+1] - bi[Crow]; 1876 cols = bj + bi[Crow]; 1877 vals = b->a + bi[Crow]; 1878 for (k=0; k<ncols; k++) { 1879 tcol = cmap[bmap[cols[k]]]; 1880 if (tcol) { 1881 subcols[i] = --tcol; 1882 subvals[i++] = vals[k]; 1883 } 1884 } 1885 #endif 1886 ierr = MatSetValues_SeqAIJ(submat,1,&row,i,subcols,subvals,INSERT_VALUES);CHKERRQ(ierr); 1887 } 1888 } 1889 } 1890 1891 /* Now assemble the off-proc rows */ 1892 for (i=0; i<nrqs; i++) { /* for each requested message */ 1893 /* recv values from other processes */ 1894 ierr = MPI_Waitany(nrqs,r_waits4,&idex,r_status4+i);CHKERRQ(ierr); 1895 proc = pa[idex]; 1896 sbuf1_i = sbuf1[proc]; 1897 /* jmax = sbuf1_i[0]; if (jmax != 1)SETERRQ1(PETSC_COMM_SELF,0,"jmax %d != 1",jmax); */ 1898 ct1 = 2 + 1; 1899 ct2 = 0; /* count of received C->j */ 1900 ct3 = 0; /* count of received C->j that will be inserted into submat */ 1901 rbuf2_i = rbuf2[idex]; /* int** received length of C->j from other processes */ 1902 rbuf3_i = rbuf3[idex]; /* int** received C->j from other processes */ 1903 rbuf4_i = rbuf4[idex]; /* scalar** received C->a from other processes */ 1904 1905 /* is_no = sbuf1_i[2*j-1]; if (is_no != 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"is_no !=0"); */ 1906 max1 = sbuf1_i[2]; /* num of rows */ 1907 for (k=0; k<max1; k++,ct1++) { /* for each recved row */ 1908 row = sbuf1_i[ct1]; /* row index of submat */ 1909 if (!allcolumns) { 1910 idex = 0; 1911 if (scall == MAT_INITIAL_MATRIX) { 1912 nnz = rbuf2_i[ct1]; /* num of C entries in this row */ 1913 for (l=0; l<nnz; l++,ct2++) { /* for each recved column */ 1914 #if defined(PETSC_USE_CTABLE) 1915 if (rbuf3_i[ct2] >= cstart && rbuf3_i[ct2] <cend) { 1916 tcol = cmap_loc[rbuf3_i[ct2] - cstart]; 1917 } else { 1918 ierr = PetscTableFind(cmap,rbuf3_i[ct2]+1,&tcol);CHKERRQ(ierr); 1919 } 1920 #else 1921 tcol = cmap[rbuf3_i[ct2]]; 1922 #endif 1923 if (tcol) { 1924 subcols[idex] = --tcol; 1925 subvals[idex++] = rbuf4_i[ct2]; 1926 1927 /* We receive an entire column of C, but a subset of it needs to be inserted into submat. 1928 For reuse, we replace received C->j with index that should be inserted to submat */ 1929 rbuf3_i[ct3++] = ct2; 1930 } 1931 } 1932 ierr = MatSetValues_SeqAIJ(submat,1,&row,idex,subcols,subvals,INSERT_VALUES);CHKERRQ(ierr); 1933 1934 } else { /* scall == MAT_REUSE_MATRIX */ 1935 submat = submats[0]; 1936 subc = (Mat_SeqAIJ*)submat->data; 1937 1938 nnz = subc->i[row+1] - subc->i[row]; /* num of submat entries in this row */ 1939 for (l=0; l<nnz; l++) { 1940 ct2 = rbuf3_i[ct3++]; /* index of rbuf4_i[] which needs to be inserted into submat */ 1941 subvals[idex++] = rbuf4_i[ct2]; 1942 } 1943 1944 bj = subc->j + subc->i[row]; 1945 ierr = MatSetValues_SeqAIJ(submat,1,&row,nnz,bj,subvals,INSERT_VALUES);CHKERRQ(ierr); 1946 } 1947 } else { /* allcolumns */ 1948 nnz = rbuf2_i[ct1]; /* num of C entries in this row */ 1949 ierr = MatSetValues_SeqAIJ(submat,1,&row,nnz,rbuf3_i+ct2,rbuf4_i+ct2,INSERT_VALUES);CHKERRQ(ierr); 1950 ct2 += nnz; 1951 } 1952 } 1953 } 1954 1955 /* sending a->a are done */ 1956 ierr = MPI_Waitall(nrqr,s_waits4,s_status4);CHKERRQ(ierr); 1957 ierr = PetscFree4(r_waits4,s_waits4,r_status4,s_status4);CHKERRQ(ierr); 1958 1959 ierr = MatAssemblyBegin(submat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1960 ierr = MatAssemblyEnd(submat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1961 submats[0] = submat; 1962 1963 /* Restore the indices */ 1964 ierr = ISRestoreIndices(isrow[0],&irow);CHKERRQ(ierr); 1965 if (!allcolumns) { 1966 ierr = ISRestoreIndices(iscol[0],&icol);CHKERRQ(ierr); 1967 } 1968 1969 /* Destroy allocated memory */ 1970 for (i=0; i<nrqs; ++i) { 1971 ierr = PetscFree3(rbuf4[i],subcols,subvals);CHKERRQ(ierr); 1972 } 1973 ierr = PetscFree3(rbuf4,subcols,subvals);CHKERRQ(ierr); 1974 ierr = PetscFree(sbuf_aa[0]);CHKERRQ(ierr); 1975 ierr = PetscFree(sbuf_aa);CHKERRQ(ierr); 1976 1977 if (scall == MAT_INITIAL_MATRIX) { 1978 ierr = PetscFree(lens);CHKERRQ(ierr); 1979 ierr = PetscFree(sbuf_aj[0]);CHKERRQ(ierr); 1980 ierr = PetscFree(sbuf_aj);CHKERRQ(ierr); 1981 } 1982 PetscFunctionReturn(0); 1983 } 1984 1985 PetscErrorCode MatGetSubMatrices_MPIAIJ_SingleIS(Mat C,PetscInt ismax,const IS isrow[],const IS iscol[],MatReuse scall,Mat *submat[]) 1986 { 1987 PetscErrorCode ierr; 1988 PetscInt ncol; 1989 PetscBool colflag,allcolumns=PETSC_FALSE; 1990 1991 PetscFunctionBegin; 1992 /* Allocate memory to hold all the submatrices */ 1993 if (scall == MAT_INITIAL_MATRIX) { 1994 ierr = PetscMalloc1(1,submat);CHKERRQ(ierr); 1995 } 1996 1997 /* Check for special case: each processor gets entire matrix columns */ 1998 ierr = ISIdentity(iscol[0],&colflag);CHKERRQ(ierr); 1999 ierr = ISGetLocalSize(iscol[0],&ncol);CHKERRQ(ierr); 2000 if (colflag && ncol == C->cmap->N) allcolumns = PETSC_TRUE; 2001 2002 ierr = MatGetSubMatrices_MPIAIJ_SingleIS_Local(C,ismax,isrow,iscol,scall,allcolumns,*submat);CHKERRQ(ierr); 2003 PetscFunctionReturn(0); 2004 } 2005 2006 PetscErrorCode MatGetSubMatrices_MPIAIJ(Mat C,PetscInt ismax,const IS isrow[],const IS iscol[],MatReuse scall,Mat *submat[]) 2007 { 2008 PetscErrorCode ierr; 2009 PetscInt nmax,nstages_local,nstages,i,pos,max_no,nrow,ncol; 2010 PetscBool rowflag,colflag,wantallmatrix=PETSC_FALSE,twantallmatrix,*allcolumns; 2011 2012 PetscFunctionBegin; 2013 #if 0 2014 /* Check for special case: each processor gets entire matrix */ 2015 if (C->submat_singleis) { /* flag is set in PCSetUp_ASM() to skip several MPIU_Allreduce() */ 2016 ierr = MatGetSubMatrices_MPIAIJ_SingleIS(C,ismax,isrow,iscol,scall,submat);CHKERRQ(ierr); 2017 PetscFunctionReturn(0); 2018 } 2019 #endif 2020 2021 if (ismax == 1 && C->rmap->N == C->cmap->N) { 2022 ierr = ISIdentity(*isrow,&rowflag);CHKERRQ(ierr); 2023 ierr = ISIdentity(*iscol,&colflag);CHKERRQ(ierr); 2024 ierr = ISGetLocalSize(*isrow,&nrow);CHKERRQ(ierr); 2025 ierr = ISGetLocalSize(*iscol,&ncol);CHKERRQ(ierr); 2026 if (rowflag && colflag && nrow == C->rmap->N && ncol == C->cmap->N) { 2027 wantallmatrix = PETSC_TRUE; 2028 2029 ierr = PetscOptionsGetBool(((PetscObject)C)->options,((PetscObject)C)->prefix,"-use_fast_submatrix",&wantallmatrix,NULL);CHKERRQ(ierr); 2030 } 2031 } 2032 ierr = MPIU_Allreduce(&wantallmatrix,&twantallmatrix,1,MPIU_BOOL,MPI_MIN,PetscObjectComm((PetscObject)C));CHKERRQ(ierr); 2033 if (twantallmatrix) { 2034 ierr = MatGetSubMatrix_MPIAIJ_All(C,MAT_GET_VALUES,scall,submat);CHKERRQ(ierr); 2035 PetscFunctionReturn(0); 2036 } 2037 2038 /* Allocate memory to hold all the submatrices */ 2039 if (scall == MAT_INITIAL_MATRIX) { 2040 ierr = PetscMalloc1(ismax+1,submat);CHKERRQ(ierr); 2041 if (!ismax) PetscFunctionReturn(0); 2042 } 2043 2044 /* Check for special case: each processor gets entire matrix columns */ 2045 ierr = PetscMalloc1(ismax+1,&allcolumns);CHKERRQ(ierr); 2046 for (i=0; i<ismax; i++) { 2047 ierr = ISIdentity(iscol[i],&colflag);CHKERRQ(ierr); 2048 ierr = ISGetLocalSize(iscol[i],&ncol);CHKERRQ(ierr); 2049 if (colflag && ncol == C->cmap->N) { 2050 allcolumns[i] = PETSC_TRUE; 2051 } else { 2052 allcolumns[i] = PETSC_FALSE; 2053 } 2054 } 2055 2056 /* Determine the number of stages through which submatrices are done */ 2057 nmax = 20*1000000 / (C->cmap->N * sizeof(PetscInt)); 2058 2059 /* 2060 Each stage will extract nmax submatrices. 2061 nmax is determined by the matrix column dimension. 2062 If the original matrix has 20M columns, only one submatrix per stage is allowed, etc. 2063 */ 2064 if (!nmax) nmax = 1; 2065 nstages_local = ismax/nmax + ((ismax % nmax) ? 1 : 0); 2066 2067 /* Make sure every processor loops through the nstages */ 2068 ierr = MPIU_Allreduce(&nstages_local,&nstages,1,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)C));CHKERRQ(ierr); 2069 2070 for (i=0,pos=0; i<nstages; i++) { 2071 if (pos+nmax <= ismax) max_no = nmax; 2072 else if (pos == ismax) max_no = 0; 2073 else max_no = ismax-pos; 2074 ierr = MatGetSubMatrices_MPIAIJ_Local(C,max_no,isrow+pos,iscol+pos,scall,allcolumns+pos,*submat+pos);CHKERRQ(ierr); 2075 pos += max_no; 2076 } 2077 2078 ierr = PetscFree(allcolumns);CHKERRQ(ierr); 2079 PetscFunctionReturn(0); 2080 } 2081 2082 /* -------------------------------------------------------------------------*/ 2083 PetscErrorCode MatGetSubMatrices_MPIAIJ_Local(Mat C,PetscInt ismax,const IS isrow[],const IS iscol[],MatReuse scall,PetscBool *allcolumns,Mat *submats) 2084 { 2085 Mat_MPIAIJ *c = (Mat_MPIAIJ*)C->data; 2086 Mat A = c->A; 2087 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data,*b = (Mat_SeqAIJ*)c->B->data,*mat; 2088 const PetscInt **icol,**irow; 2089 PetscInt *nrow,*ncol,start; 2090 PetscErrorCode ierr; 2091 PetscMPIInt rank,size,tag0,tag1,tag2,tag3,*w1,*w2,*w3,*w4,nrqr; 2092 PetscInt **sbuf1,**sbuf2,i,j,k,l,ct1,ct2,**rbuf1,row,proc; 2093 PetscInt nrqs,msz,**ptr,*req_size,*ctr,*pa,*tmp,tcol; 2094 PetscInt **rbuf3,*req_source,**sbuf_aj,**rbuf2,max1,max2; 2095 PetscInt **lens,is_no,ncols,*cols,mat_i,*mat_j,tmp2,jmax; 2096 #if defined(PETSC_USE_CTABLE) 2097 PetscTable *cmap,cmap_i=NULL,*rmap,rmap_i; 2098 #else 2099 PetscInt **cmap,*cmap_i=NULL,**rmap,*rmap_i; 2100 #endif 2101 const PetscInt *irow_i; 2102 PetscInt ctr_j,*sbuf1_j,*sbuf_aj_i,*rbuf1_i,kmax,*lens_i; 2103 MPI_Request *s_waits1,*r_waits1,*s_waits2,*r_waits2,*r_waits3; 2104 MPI_Request *r_waits4,*s_waits3,*s_waits4; 2105 MPI_Status *r_status1,*r_status2,*s_status1,*s_status3,*s_status2; 2106 MPI_Status *r_status3,*r_status4,*s_status4; 2107 MPI_Comm comm; 2108 PetscScalar **rbuf4,**sbuf_aa,*vals,*mat_a,*sbuf_aa_i; 2109 PetscMPIInt *onodes1,*olengths1; 2110 PetscMPIInt idex,idex2,end; 2111 PetscInt **row2proc,*row2proc_i; 2112 Mat_SubMat **smats,*smat_i; 2113 PetscBool *issorted; 2114 2115 PetscFunctionBegin; 2116 ierr = PetscObjectGetComm((PetscObject)C,&comm);CHKERRQ(ierr); 2117 tag0 = ((PetscObject)C)->tag; 2118 size = c->size; 2119 rank = c->rank; 2120 //printf("[%d] MatGetSubMatrices_MPIAIJ_Local, scall %d\n",rank,scall); 2121 2122 /* Get some new tags to keep the communication clean */ 2123 ierr = PetscObjectGetNewTag((PetscObject)C,&tag1);CHKERRQ(ierr); 2124 ierr = PetscObjectGetNewTag((PetscObject)C,&tag2);CHKERRQ(ierr); 2125 ierr = PetscObjectGetNewTag((PetscObject)C,&tag3);CHKERRQ(ierr); 2126 2127 ierr = PetscMalloc4(ismax,&irow,ismax,&icol,ismax,&nrow,ismax,&ncol);CHKERRQ(ierr); 2128 ierr = PetscMalloc1(ismax,&issorted);CHKERRQ(ierr); 2129 2130 for (i=0; i<ismax; i++) { 2131 ierr = ISSorted(isrow[i],&issorted[i]);CHKERRQ(ierr); 2132 2133 ierr = ISGetIndices(isrow[i],&irow[i]);CHKERRQ(ierr); 2134 ierr = ISGetLocalSize(isrow[i],&nrow[i]);CHKERRQ(ierr); 2135 if (allcolumns[i]) { 2136 icol[i] = NULL; 2137 ncol[i] = C->cmap->N; 2138 } else { 2139 ierr = ISGetIndices(iscol[i],&icol[i]);CHKERRQ(ierr); 2140 ierr = ISGetLocalSize(iscol[i],&ncol[i]);CHKERRQ(ierr); 2141 } 2142 } 2143 2144 2145 ierr = PetscMalloc1(ismax,&smats);CHKERRQ(ierr); 2146 ierr = PetscMalloc1(ismax,&row2proc);CHKERRQ(ierr); 2147 2148 if (scall == MAT_REUSE_MATRIX) { 2149 //PetscInt *req_source1,*req_source2; 2150 2151 /* Assumes new rows are same length as the old rows,hence bug! */ 2152 for (i=0; i<ismax; i++) { 2153 mat = (Mat_SeqAIJ*)(submats[i]->data); 2154 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"); 2155 2156 /* Initial matrix as if empty -- rm later!!! */ 2157 ierr = PetscMemzero(mat->ilen,submats[i]->rmap->n*sizeof(PetscInt));CHKERRQ(ierr); 2158 2159 /* Initial matrix as if empty */ 2160 submats[i]->factortype = C->factortype; 2161 2162 smat_i = mat->submatis1; 2163 smats[i] = smat_i; 2164 2165 nrqs = smat_i->nrqs; 2166 nrqr = smat_i->nrqr; 2167 rbuf1 = smat_i->rbuf1; 2168 rbuf2 = smat_i->rbuf2; 2169 rbuf3 = smat_i->rbuf3; 2170 //req_source2 = smat_i->req_source2; 2171 2172 sbuf1 = smat_i->sbuf1; 2173 sbuf2 = smat_i->sbuf2; 2174 ptr = smat_i->ptr; 2175 tmp = smat_i->tmp; 2176 ctr = smat_i->ctr; 2177 2178 pa = smat_i->pa; 2179 req_size = smat_i->req_size; 2180 //req_source1 = smat_i->req_source1; 2181 2182 allcolumns[i] = smat_i->allcolumns; 2183 row2proc[i] = smat_i->row2proc; 2184 //rmap = smat_i->rmap; 2185 //cmap = smat_i->cmap; 2186 #if defined(PETSC_USE_CTABLE) 2187 //rmap_loc = smat_i->rmap_loc; 2188 //cmap_loc = smat_i->cmap_loc; 2189 #endif 2190 } 2191 } 2192 2193 /* evaluate communication - mesg to who, length of mesg, and buffer space 2194 required. Based on this, buffers are allocated, and data copied into them*/ 2195 ierr = PetscMalloc4(size,&w1,size,&w2,size,&w3,size,&w4);CHKERRQ(ierr); /* mesg size */ 2196 ierr = PetscMemzero(w1,size*sizeof(PetscMPIInt));CHKERRQ(ierr); /* initialize work vector*/ 2197 ierr = PetscMemzero(w2,size*sizeof(PetscMPIInt));CHKERRQ(ierr); /* initialize work vector*/ 2198 ierr = PetscMemzero(w3,size*sizeof(PetscMPIInt));CHKERRQ(ierr); /* initialize work vector*/ 2199 2200 for (i=0; i<ismax; i++) { 2201 ierr = PetscMemzero(w4,size*sizeof(PetscMPIInt));CHKERRQ(ierr); /* initialize work vector*/ 2202 jmax = nrow[i]; 2203 irow_i = irow[i]; 2204 2205 if (scall == MAT_INITIAL_MATRIX) { 2206 ierr = PetscMalloc1(jmax,&row2proc_i);CHKERRQ(ierr); 2207 row2proc[i] = row2proc_i; 2208 } else { 2209 row2proc_i = row2proc[i]; 2210 } 2211 2212 if (issorted[i]) proc = 0; 2213 for (j=0; j<jmax; j++) { 2214 if (!issorted[i]) proc = 0; 2215 row = irow_i[j]; 2216 while (row >= C->rmap->range[proc+1]) proc++; 2217 w4[proc]++; 2218 row2proc_i[j] = proc; /* map row index to proc */ 2219 } 2220 for (j=0; j<size; j++) { 2221 if (w4[j]) { w1[j] += w4[j]; w3[j]++;} 2222 } 2223 } 2224 2225 nrqs = 0; /* no of outgoing messages */ 2226 msz = 0; /* total mesg length (for all procs) */ 2227 w1[rank] = 0; /* no mesg sent to self */ 2228 w3[rank] = 0; 2229 for (i=0; i<size; i++) { 2230 if (w1[i]) { w2[i] = 1; nrqs++;} /* there exists a message to proc i */ 2231 } 2232 ierr = PetscMalloc1(nrqs+1,&pa);CHKERRQ(ierr); /*(proc -array)*/ 2233 for (i=0,j=0; i<size; i++) { 2234 if (w1[i]) { pa[j] = i; j++; } 2235 } 2236 2237 /* Each message would have a header = 1 + 2*(no of IS) + data */ 2238 for (i=0; i<nrqs; i++) { 2239 j = pa[i]; 2240 w1[j] += w2[j] + 2* w3[j]; 2241 msz += w1[j]; 2242 } 2243 ierr = PetscInfo2(0,"Number of outgoing messages %D Total message length %D\n",nrqs,msz);CHKERRQ(ierr); 2244 2245 /* Determine the number of messages to expect, their lengths, from from-ids */ 2246 ierr = PetscGatherNumberOfMessages(comm,w2,w1,&nrqr);CHKERRQ(ierr); 2247 ierr = PetscGatherMessageLengths(comm,nrqs,nrqr,w1,&onodes1,&olengths1);CHKERRQ(ierr); 2248 2249 /* Now post the Irecvs corresponding to these messages */ 2250 ierr = PetscPostIrecvInt(comm,tag0,nrqr,onodes1,olengths1,&rbuf1,&r_waits1);CHKERRQ(ierr); 2251 2252 ierr = PetscFree(onodes1);CHKERRQ(ierr); 2253 ierr = PetscFree(olengths1);CHKERRQ(ierr); 2254 2255 /* Allocate Memory for outgoing messages */ 2256 ierr = PetscMalloc4(size,&sbuf1,size,&ptr,2*msz,&tmp,size,&ctr);CHKERRQ(ierr); 2257 ierr = PetscMemzero(sbuf1,size*sizeof(PetscInt*));CHKERRQ(ierr); 2258 ierr = PetscMemzero(ptr,size*sizeof(PetscInt*));CHKERRQ(ierr); 2259 2260 { 2261 PetscInt *iptr = tmp; 2262 k = 0; 2263 for (i=0; i<nrqs; i++) { 2264 j = pa[i]; 2265 iptr += k; 2266 sbuf1[j] = iptr; 2267 k = w1[j]; 2268 } 2269 } 2270 2271 /* Form the outgoing messages */ 2272 /* Initialize the header space */ 2273 for (i=0; i<nrqs; i++) { 2274 j = pa[i]; 2275 sbuf1[j][0] = 0; 2276 ierr = PetscMemzero(sbuf1[j]+1,2*w3[j]*sizeof(PetscInt));CHKERRQ(ierr); 2277 ptr[j] = sbuf1[j] + 2*w3[j] + 1; 2278 } 2279 2280 /* Parse the isrow and copy data into outbuf */ 2281 for (i=0; i<ismax; i++) { 2282 row2proc_i = row2proc[i]; 2283 ierr = PetscMemzero(ctr,size*sizeof(PetscInt));CHKERRQ(ierr); 2284 irow_i = irow[i]; 2285 jmax = nrow[i]; 2286 for (j=0; j<jmax; j++) { /* parse the indices of each IS */ 2287 proc = row2proc_i[j]; 2288 if (proc != rank) { /* copy to the outgoing buf*/ 2289 ctr[proc]++; 2290 *ptr[proc] = irow_i[j]; 2291 ptr[proc]++; 2292 } 2293 } 2294 /* Update the headers for the current IS */ 2295 for (j=0; j<size; j++) { /* Can Optimise this loop too */ 2296 if ((ctr_j = ctr[j])) { 2297 sbuf1_j = sbuf1[j]; 2298 k = ++sbuf1_j[0]; 2299 sbuf1_j[2*k] = ctr_j; 2300 sbuf1_j[2*k-1] = i; 2301 } 2302 } 2303 } 2304 2305 /* Now post the sends */ 2306 ierr = PetscMalloc1(nrqs+1,&s_waits1);CHKERRQ(ierr); 2307 for (i=0; i<nrqs; ++i) { 2308 j = pa[i]; 2309 ierr = MPI_Isend(sbuf1[j],w1[j],MPIU_INT,j,tag0,comm,s_waits1+i);CHKERRQ(ierr); 2310 } 2311 2312 /* Post Receives to capture the buffer size */ 2313 ierr = PetscMalloc1(nrqs+1,&r_waits2);CHKERRQ(ierr); 2314 ierr = PetscMalloc1(nrqs+1,&rbuf2);CHKERRQ(ierr); 2315 rbuf2[0] = tmp + msz; 2316 for (i=1; i<nrqs; ++i) { 2317 rbuf2[i] = rbuf2[i-1]+w1[pa[i-1]]; 2318 } 2319 for (i=0; i<nrqs; ++i) { 2320 j = pa[i]; 2321 ierr = MPI_Irecv(rbuf2[i],w1[j],MPIU_INT,j,tag1,comm,r_waits2+i);CHKERRQ(ierr); 2322 } 2323 2324 /* Send to other procs the buf size they should allocate */ 2325 2326 2327 /* Receive messages*/ 2328 ierr = PetscMalloc1(nrqr+1,&s_waits2);CHKERRQ(ierr); 2329 ierr = PetscMalloc1(nrqr+1,&r_status1);CHKERRQ(ierr); 2330 ierr = PetscMalloc3(nrqr,&sbuf2,nrqr,&req_size,nrqr,&req_source);CHKERRQ(ierr); 2331 { 2332 PetscInt *sAi = a->i,*sBi = b->i,id,rstart = C->rmap->rstart; 2333 PetscInt *sbuf2_i; 2334 2335 for (i=0; i<nrqr; ++i) { 2336 ierr = MPI_Waitany(nrqr,r_waits1,&idex,r_status1+i);CHKERRQ(ierr); 2337 2338 req_size[idex] = 0; 2339 rbuf1_i = rbuf1[idex]; 2340 start = 2*rbuf1_i[0] + 1; 2341 ierr = MPI_Get_count(r_status1+i,MPIU_INT,&end);CHKERRQ(ierr); 2342 ierr = PetscMalloc1(end+1,&sbuf2[idex]);CHKERRQ(ierr); 2343 sbuf2_i = sbuf2[idex]; 2344 for (j=start; j<end; j++) { 2345 id = rbuf1_i[j] - rstart; 2346 ncols = sAi[id+1] - sAi[id] + sBi[id+1] - sBi[id]; 2347 sbuf2_i[j] = ncols; 2348 req_size[idex] += ncols; 2349 } 2350 req_source[idex] = r_status1[i].MPI_SOURCE; 2351 /* form the header */ 2352 sbuf2_i[0] = req_size[idex]; 2353 for (j=1; j<start; j++) sbuf2_i[j] = rbuf1_i[j]; 2354 2355 ierr = MPI_Isend(sbuf2_i,end,MPIU_INT,req_source[idex],tag1,comm,s_waits2+i);CHKERRQ(ierr); 2356 } 2357 } 2358 ierr = PetscFree(r_status1);CHKERRQ(ierr); 2359 ierr = PetscFree(r_waits1);CHKERRQ(ierr); 2360 2361 /* recv buffer sizes */ 2362 /* Receive messages*/ 2363 2364 ierr = PetscMalloc1(nrqs+1,&rbuf3);CHKERRQ(ierr); 2365 ierr = PetscMalloc1(nrqs+1,&rbuf4);CHKERRQ(ierr); 2366 ierr = PetscMalloc1(nrqs+1,&r_waits3);CHKERRQ(ierr); 2367 ierr = PetscMalloc1(nrqs+1,&r_waits4);CHKERRQ(ierr); 2368 ierr = PetscMalloc1(nrqs+1,&r_status2);CHKERRQ(ierr); 2369 2370 for (i=0; i<nrqs; ++i) { 2371 ierr = MPI_Waitany(nrqs,r_waits2,&idex,r_status2+i);CHKERRQ(ierr); 2372 ierr = PetscMalloc1(rbuf2[idex][0]+1,&rbuf3[idex]);CHKERRQ(ierr); 2373 ierr = PetscMalloc1(rbuf2[idex][0]+1,&rbuf4[idex]);CHKERRQ(ierr); 2374 ierr = MPI_Irecv(rbuf3[idex],rbuf2[idex][0],MPIU_INT,r_status2[i].MPI_SOURCE,tag2,comm,r_waits3+idex);CHKERRQ(ierr); 2375 ierr = MPI_Irecv(rbuf4[idex],rbuf2[idex][0],MPIU_SCALAR,r_status2[i].MPI_SOURCE,tag3,comm,r_waits4+idex);CHKERRQ(ierr); 2376 } 2377 ierr = PetscFree(r_status2);CHKERRQ(ierr); 2378 ierr = PetscFree(r_waits2);CHKERRQ(ierr); 2379 2380 /* Wait on sends1 and sends2 */ 2381 ierr = PetscMalloc1(nrqs+1,&s_status1);CHKERRQ(ierr); 2382 ierr = PetscMalloc1(nrqr+1,&s_status2);CHKERRQ(ierr); 2383 2384 if (nrqs) {ierr = MPI_Waitall(nrqs,s_waits1,s_status1);CHKERRQ(ierr);} 2385 if (nrqr) {ierr = MPI_Waitall(nrqr,s_waits2,s_status2);CHKERRQ(ierr);} 2386 ierr = PetscFree(s_status1);CHKERRQ(ierr); 2387 ierr = PetscFree(s_status2);CHKERRQ(ierr); 2388 ierr = PetscFree(s_waits1);CHKERRQ(ierr); 2389 ierr = PetscFree(s_waits2);CHKERRQ(ierr); 2390 2391 /* Now allocate buffers for a->j, and send them off */ 2392 ierr = PetscMalloc1(nrqr+1,&sbuf_aj);CHKERRQ(ierr); 2393 for (i=0,j=0; i<nrqr; i++) j += req_size[i]; 2394 ierr = PetscMalloc1(j+1,&sbuf_aj[0]);CHKERRQ(ierr); 2395 for (i=1; i<nrqr; i++) sbuf_aj[i] = sbuf_aj[i-1] + req_size[i-1]; 2396 2397 ierr = PetscMalloc1(nrqr+1,&s_waits3);CHKERRQ(ierr); 2398 { 2399 PetscInt nzA,nzB,*a_i = a->i,*b_i = b->i,lwrite; 2400 PetscInt *cworkA,*cworkB,cstart = C->cmap->rstart,rstart = C->rmap->rstart,*bmap = c->garray; 2401 PetscInt cend = C->cmap->rend; 2402 PetscInt *a_j = a->j,*b_j = b->j,ctmp; 2403 2404 for (i=0; i<nrqr; i++) { 2405 rbuf1_i = rbuf1[i]; 2406 sbuf_aj_i = sbuf_aj[i]; 2407 ct1 = 2*rbuf1_i[0] + 1; 2408 ct2 = 0; 2409 for (j=1,max1=rbuf1_i[0]; j<=max1; j++) { 2410 kmax = rbuf1[i][2*j]; 2411 for (k=0; k<kmax; k++,ct1++) { 2412 row = rbuf1_i[ct1] - rstart; 2413 nzA = a_i[row+1] - a_i[row]; nzB = b_i[row+1] - b_i[row]; 2414 ncols = nzA + nzB; 2415 cworkA = a_j + a_i[row]; cworkB = b_j + b_i[row]; 2416 2417 /* load the column indices for this row into cols*/ 2418 cols = sbuf_aj_i + ct2; 2419 2420 lwrite = 0; 2421 for (l=0; l<nzB; l++) { 2422 if ((ctmp = bmap[cworkB[l]]) < cstart) cols[lwrite++] = ctmp; 2423 } 2424 for (l=0; l<nzA; l++) cols[lwrite++] = cstart + cworkA[l]; 2425 for (l=0; l<nzB; l++) { 2426 if ((ctmp = bmap[cworkB[l]]) >= cend) cols[lwrite++] = ctmp; 2427 } 2428 2429 ct2 += ncols; 2430 } 2431 } 2432 ierr = MPI_Isend(sbuf_aj_i,req_size[i],MPIU_INT,req_source[i],tag2,comm,s_waits3+i);CHKERRQ(ierr); 2433 } 2434 } 2435 ierr = PetscMalloc1(nrqs+1,&r_status3);CHKERRQ(ierr); 2436 ierr = PetscMalloc1(nrqr+1,&s_status3);CHKERRQ(ierr); 2437 2438 ierr = PetscMalloc1(nrqs+1,&r_status4);CHKERRQ(ierr); 2439 ierr = PetscMalloc1(nrqr+1,&s_status4);CHKERRQ(ierr); 2440 2441 /* Form the matrix */ 2442 /* create col map: global col of C -> local col of submatrices */ 2443 { 2444 const PetscInt *icol_i; 2445 #if defined(PETSC_USE_CTABLE) 2446 ierr = PetscMalloc1(1+ismax,&cmap);CHKERRQ(ierr); 2447 for (i=0; i<ismax; i++) { 2448 if (!allcolumns[i]) { 2449 ierr = PetscTableCreate(ncol[i]+1,C->cmap->N+1,&cmap[i]);CHKERRQ(ierr); 2450 2451 jmax = ncol[i]; 2452 icol_i = icol[i]; 2453 cmap_i = cmap[i]; 2454 for (j=0; j<jmax; j++) { 2455 ierr = PetscTableAdd(cmap[i],icol_i[j]+1,j+1,INSERT_VALUES);CHKERRQ(ierr); 2456 } 2457 } else { 2458 cmap[i] = NULL; 2459 } 2460 } 2461 #else 2462 ierr = PetscMalloc1(ismax,&cmap);CHKERRQ(ierr); 2463 for (i=0; i<ismax; i++) { 2464 if (!allcolumns[i]) { 2465 ierr = PetscMalloc1(C->cmap->N,&cmap[i]);CHKERRQ(ierr); 2466 ierr = PetscMemzero(cmap[i],C->cmap->N*sizeof(PetscInt));CHKERRQ(ierr); 2467 jmax = ncol[i]; 2468 icol_i = icol[i]; 2469 cmap_i = cmap[i]; 2470 for (j=0; j<jmax; j++) { 2471 cmap_i[icol_i[j]] = j+1; 2472 } 2473 } else { 2474 cmap[i] = NULL; 2475 } 2476 } 2477 #endif 2478 } 2479 2480 /* Create lens which is required for MatCreate... */ 2481 for (i=0,j=0; i<ismax; i++) j += nrow[i]; 2482 ierr = PetscMalloc1(ismax,&lens);CHKERRQ(ierr); 2483 if (ismax) { 2484 ierr = PetscMalloc1(j,&lens[0]);CHKERRQ(ierr); 2485 ierr = PetscMemzero(lens[0],j*sizeof(PetscInt));CHKERRQ(ierr); 2486 } 2487 for (i=1; i<ismax; i++) lens[i] = lens[i-1] + nrow[i-1]; 2488 2489 /* Update lens from local data */ 2490 for (i=0; i<ismax; i++) { 2491 row2proc_i = row2proc[i]; 2492 jmax = nrow[i]; 2493 if (!allcolumns[i]) cmap_i = cmap[i]; 2494 irow_i = irow[i]; 2495 lens_i = lens[i]; 2496 for (j=0; j<jmax; j++) { 2497 row = irow_i[j]; 2498 proc = row2proc_i[j]; 2499 if (proc == rank) { 2500 ierr = MatGetRow_MPIAIJ(C,row,&ncols,&cols,0);CHKERRQ(ierr); 2501 if (!allcolumns[i]) { 2502 for (k=0; k<ncols; k++) { 2503 #if defined(PETSC_USE_CTABLE) 2504 ierr = PetscTableFind(cmap_i,cols[k]+1,&tcol);CHKERRQ(ierr); 2505 #else 2506 tcol = cmap_i[cols[k]]; 2507 #endif 2508 if (tcol) lens_i[j]++; 2509 } 2510 } else { /* allcolumns */ 2511 lens_i[j] = ncols; 2512 } 2513 ierr = MatRestoreRow_MPIAIJ(C,row,&ncols,&cols,0);CHKERRQ(ierr); 2514 } 2515 } 2516 } 2517 2518 /* Create row map: global row of C -> local row of submatrices */ 2519 #if defined(PETSC_USE_CTABLE) 2520 ierr = PetscMalloc1(1+ismax,&rmap);CHKERRQ(ierr); 2521 for (i=0; i<ismax; i++) { 2522 ierr = PetscTableCreate(nrow[i]+1,C->rmap->N+1,&rmap[i]);CHKERRQ(ierr); 2523 irow_i = irow[i]; 2524 jmax = nrow[i]; 2525 for (j=0; j<jmax; j++) { 2526 ierr = PetscTableAdd(rmap[i],irow_i[j]+1,j+1,INSERT_VALUES);CHKERRQ(ierr); 2527 } 2528 } 2529 #else 2530 ierr = PetscMalloc1(ismax,&rmap);CHKERRQ(ierr); 2531 if (ismax) { 2532 ierr = PetscMalloc1(ismax*C->rmap->N,&rmap[0]);CHKERRQ(ierr); 2533 ierr = PetscMemzero(rmap[0],ismax*C->rmap->N*sizeof(PetscInt));CHKERRQ(ierr); 2534 } 2535 for (i=1; i<ismax; i++) rmap[i] = rmap[i-1] + C->rmap->N; 2536 for (i=0; i<ismax; i++) { 2537 rmap_i = rmap[i]; 2538 irow_i = irow[i]; 2539 jmax = nrow[i]; 2540 for (j=0; j<jmax; j++) { 2541 rmap_i[irow_i[j]] = j; 2542 } 2543 } 2544 #endif 2545 2546 /* Update lens from offproc data */ 2547 { 2548 PetscInt *rbuf2_i,*rbuf3_i,*sbuf1_i; 2549 2550 for (tmp2=0; tmp2<nrqs; tmp2++) { 2551 ierr = MPI_Waitany(nrqs,r_waits3,&idex2,r_status3+tmp2);CHKERRQ(ierr); 2552 idex = pa[idex2]; 2553 sbuf1_i = sbuf1[idex]; 2554 jmax = sbuf1_i[0]; 2555 ct1 = 2*jmax+1; 2556 ct2 = 0; 2557 rbuf2_i = rbuf2[idex2]; 2558 rbuf3_i = rbuf3[idex2]; 2559 for (j=1; j<=jmax; j++) { 2560 is_no = sbuf1_i[2*j-1]; 2561 max1 = sbuf1_i[2*j]; 2562 lens_i = lens[is_no]; 2563 if (!allcolumns[is_no]) cmap_i = cmap[is_no]; 2564 rmap_i = rmap[is_no]; 2565 for (k=0; k<max1; k++,ct1++) { 2566 #if defined(PETSC_USE_CTABLE) 2567 ierr = PetscTableFind(rmap_i,sbuf1_i[ct1]+1,&row);CHKERRQ(ierr); 2568 row--; 2569 if (row < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"row not found in table"); 2570 #else 2571 row = rmap_i[sbuf1_i[ct1]]; /* the val in the new matrix to be */ 2572 #endif 2573 max2 = rbuf2_i[ct1]; 2574 for (l=0; l<max2; l++,ct2++) { 2575 if (!allcolumns[is_no]) { 2576 #if defined(PETSC_USE_CTABLE) 2577 ierr = PetscTableFind(cmap_i,rbuf3_i[ct2]+1,&tcol);CHKERRQ(ierr); 2578 #else 2579 tcol = cmap_i[rbuf3_i[ct2]]; 2580 #endif 2581 if (tcol) lens_i[row]++; 2582 } else { /* allcolumns */ 2583 lens_i[row]++; /* lens_i[row] += max2 ? */ 2584 } 2585 } 2586 } 2587 } 2588 } 2589 } 2590 ierr = PetscFree(r_status3);CHKERRQ(ierr); 2591 ierr = PetscFree(r_waits3);CHKERRQ(ierr); 2592 if (nrqr) {ierr = MPI_Waitall(nrqr,s_waits3,s_status3);CHKERRQ(ierr);} 2593 ierr = PetscFree(s_status3);CHKERRQ(ierr); 2594 ierr = PetscFree(s_waits3);CHKERRQ(ierr); 2595 2596 /* Create the submatrices */ 2597 if (scall == MAT_INITIAL_MATRIX) { 2598 for (i=0; i<ismax; i++) { 2599 PetscInt rbs,cbs; 2600 Mat_SeqAIJ* subc; 2601 2602 ierr = ISGetBlockSize(isrow[i],&rbs);CHKERRQ(ierr); 2603 ierr = ISGetBlockSize(iscol[i],&cbs);CHKERRQ(ierr); 2604 2605 ierr = MatCreate(PETSC_COMM_SELF,submats+i);CHKERRQ(ierr); 2606 ierr = MatSetSizes(submats[i],nrow[i],ncol[i],PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 2607 2608 ierr = MatSetBlockSizes(submats[i],rbs,cbs);CHKERRQ(ierr); 2609 ierr = MatSetType(submats[i],((PetscObject)A)->type_name);CHKERRQ(ierr); 2610 ierr = MatSeqAIJSetPreallocation(submats[i],0,lens[i]);CHKERRQ(ierr); 2611 2612 /* create struct Mat_SubMat and attached it to submat */ 2613 ierr = PetscNew(&smat_i);CHKERRQ(ierr); 2614 subc = (Mat_SeqAIJ*)submats[i]->data; 2615 subc->submatis1 = smat_i; 2616 smats[i] = smat_i; 2617 2618 smat_i->destroy = submats[i]->ops->destroy; 2619 submats[i]->ops->destroy = MatDestroy_MPIAIJ_MatGetSubmatrices; 2620 submats[i]->factortype = C->factortype; 2621 2622 smat_i->nrqs = 0;//nrqs; 2623 smat_i->nrqr = 0;//nrqr; 2624 smat_i->rbuf1 = NULL;//rbuf1; 2625 smat_i->rbuf2 = NULL;//rbuf2; 2626 smat_i->rbuf3 = NULL;//rbuf3; 2627 smat_i->sbuf2 = NULL;//sbuf2; 2628 smat_i->req_source2 = NULL;//req_source2; 2629 2630 smat_i->sbuf1 = NULL;//sbuf1; 2631 smat_i->ptr = NULL;//ptr; 2632 smat_i->tmp = NULL;//tmp; 2633 smat_i->ctr = NULL;//ctr; 2634 2635 smat_i->pa = NULL;//pa; 2636 smat_i->req_size = NULL;//req_size; 2637 smat_i->req_source1 = NULL;//req_source1; 2638 2639 smat_i->allcolumns = allcolumns[i]; 2640 smat_i->row2proc = row2proc[i]; 2641 smat_i->rmap = NULL;//rmap; 2642 smat_i->cmap = NULL;//cmap; 2643 #if defined(PETSC_USE_CTABLE) 2644 smat_i->rmap_loc = NULL;//rmap_loc; 2645 smat_i->cmap_loc = NULL;//cmap_loc; 2646 #endif 2647 2648 } 2649 } 2650 2651 /* Allocate buffers for a->a, and send them off */ 2652 ierr = PetscMalloc1(nrqr+1,&sbuf_aa);CHKERRQ(ierr); 2653 for (i=0,j=0; i<nrqr; i++) j += req_size[i]; 2654 ierr = PetscMalloc1(j+1,&sbuf_aa[0]);CHKERRQ(ierr); 2655 for (i=1; i<nrqr; i++) sbuf_aa[i] = sbuf_aa[i-1] + req_size[i-1]; 2656 2657 ierr = PetscMalloc1(nrqr+1,&s_waits4);CHKERRQ(ierr); 2658 { 2659 PetscInt nzA,nzB,*a_i = a->i,*b_i = b->i, *cworkB,lwrite; 2660 PetscInt cstart = C->cmap->rstart,rstart = C->rmap->rstart,*bmap = c->garray; 2661 PetscInt cend = C->cmap->rend; 2662 PetscInt *b_j = b->j; 2663 PetscScalar *vworkA,*vworkB,*a_a = a->a,*b_a = b->a; 2664 2665 for (i=0; i<nrqr; i++) { 2666 rbuf1_i = rbuf1[i]; 2667 sbuf_aa_i = sbuf_aa[i]; 2668 ct1 = 2*rbuf1_i[0]+1; 2669 ct2 = 0; 2670 for (j=1,max1=rbuf1_i[0]; j<=max1; j++) { 2671 kmax = rbuf1_i[2*j]; 2672 for (k=0; k<kmax; k++,ct1++) { 2673 row = rbuf1_i[ct1] - rstart; 2674 nzA = a_i[row+1] - a_i[row]; nzB = b_i[row+1] - b_i[row]; 2675 ncols = nzA + nzB; 2676 cworkB = b_j + b_i[row]; 2677 vworkA = a_a + a_i[row]; 2678 vworkB = b_a + b_i[row]; 2679 2680 /* load the column values for this row into vals*/ 2681 vals = sbuf_aa_i+ct2; 2682 2683 lwrite = 0; 2684 for (l=0; l<nzB; l++) { 2685 if ((bmap[cworkB[l]]) < cstart) vals[lwrite++] = vworkB[l]; 2686 } 2687 for (l=0; l<nzA; l++) vals[lwrite++] = vworkA[l]; 2688 for (l=0; l<nzB; l++) { 2689 if ((bmap[cworkB[l]]) >= cend) vals[lwrite++] = vworkB[l]; 2690 } 2691 2692 ct2 += ncols; 2693 } 2694 } 2695 ierr = MPI_Isend(sbuf_aa_i,req_size[i],MPIU_SCALAR,req_source[i],tag3,comm,s_waits4+i);CHKERRQ(ierr); 2696 } 2697 } 2698 ierr = PetscFree(rbuf1[0]);CHKERRQ(ierr); 2699 ierr = PetscFree(rbuf1);CHKERRQ(ierr); 2700 2701 2702 /* Assemble the matrices */ 2703 /* First assemble the local rows */ 2704 { 2705 PetscInt ilen_row,*imat_ilen,*imat_j,*imat_i,old_row; 2706 PetscScalar *imat_a; 2707 2708 for (i=0; i<ismax; i++) { 2709 row2proc_i = row2proc[i]; 2710 mat = (Mat_SeqAIJ*)submats[i]->data; 2711 imat_ilen = mat->ilen; 2712 imat_j = mat->j; 2713 imat_i = mat->i; 2714 imat_a = mat->a; 2715 2716 if (!allcolumns[i]) cmap_i = cmap[i]; 2717 rmap_i = rmap[i]; 2718 irow_i = irow[i]; 2719 jmax = nrow[i]; 2720 for (j=0; j<jmax; j++) { 2721 row = irow_i[j]; 2722 proc = row2proc_i[j]; 2723 if (proc == rank) { 2724 old_row = row; 2725 #if defined(PETSC_USE_CTABLE) 2726 ierr = PetscTableFind(rmap_i,row+1,&row);CHKERRQ(ierr); 2727 row--; 2728 #else 2729 row = rmap_i[row]; 2730 #endif 2731 ilen_row = imat_ilen[row]; 2732 ierr = MatGetRow_MPIAIJ(C,old_row,&ncols,&cols,&vals);CHKERRQ(ierr); 2733 mat_i = imat_i[row]; 2734 mat_a = imat_a + mat_i; 2735 mat_j = imat_j + mat_i; 2736 if (!allcolumns[i]) { 2737 for (k=0; k<ncols; k++) { 2738 #if defined(PETSC_USE_CTABLE) 2739 ierr = PetscTableFind(cmap_i,cols[k]+1,&tcol);CHKERRQ(ierr); 2740 #else 2741 tcol = cmap_i[cols[k]]; 2742 #endif 2743 if (tcol) { 2744 *mat_j++ = tcol - 1; 2745 *mat_a++ = vals[k]; 2746 ilen_row++; 2747 } 2748 } 2749 } else { /* allcolumns */ 2750 for (k=0; k<ncols; k++) { 2751 *mat_j++ = cols[k]; /* global col index! */ 2752 *mat_a++ = vals[k]; 2753 ilen_row++; 2754 } 2755 } 2756 ierr = MatRestoreRow_MPIAIJ(C,old_row,&ncols,&cols,&vals);CHKERRQ(ierr); 2757 2758 imat_ilen[row] = ilen_row; 2759 } 2760 } 2761 } 2762 } 2763 2764 /* Now assemble the off proc rows*/ 2765 { 2766 PetscInt *sbuf1_i,*rbuf2_i,*rbuf3_i,*imat_ilen,ilen; 2767 PetscInt *imat_j,*imat_i; 2768 PetscScalar *imat_a,*rbuf4_i; 2769 2770 for (tmp2=0; tmp2<nrqs; tmp2++) { 2771 ierr = MPI_Waitany(nrqs,r_waits4,&idex2,r_status4+tmp2);CHKERRQ(ierr); 2772 idex = pa[idex2]; 2773 sbuf1_i = sbuf1[idex]; 2774 jmax = sbuf1_i[0]; 2775 ct1 = 2*jmax + 1; 2776 ct2 = 0; 2777 rbuf2_i = rbuf2[idex2]; 2778 rbuf3_i = rbuf3[idex2]; 2779 rbuf4_i = rbuf4[idex2]; 2780 for (j=1; j<=jmax; j++) { 2781 is_no = sbuf1_i[2*j-1]; 2782 rmap_i = rmap[is_no]; 2783 if (!allcolumns[is_no]) cmap_i = cmap[is_no]; 2784 mat = (Mat_SeqAIJ*)submats[is_no]->data; 2785 imat_ilen = mat->ilen; 2786 imat_j = mat->j; 2787 imat_i = mat->i; 2788 imat_a = mat->a; 2789 max1 = sbuf1_i[2*j]; 2790 for (k=0; k<max1; k++,ct1++) { 2791 row = sbuf1_i[ct1]; 2792 #if defined(PETSC_USE_CTABLE) 2793 ierr = PetscTableFind(rmap_i,row+1,&row);CHKERRQ(ierr); 2794 row--; 2795 #else 2796 row = rmap_i[row]; 2797 #endif 2798 ilen = imat_ilen[row]; 2799 mat_i = imat_i[row]; 2800 mat_a = imat_a + mat_i; 2801 mat_j = imat_j + mat_i; 2802 max2 = rbuf2_i[ct1]; 2803 if (!allcolumns[is_no]) { 2804 for (l=0; l<max2; l++,ct2++) { 2805 2806 #if defined(PETSC_USE_CTABLE) 2807 ierr = PetscTableFind(cmap_i,rbuf3_i[ct2]+1,&tcol);CHKERRQ(ierr); 2808 #else 2809 tcol = cmap_i[rbuf3_i[ct2]]; 2810 #endif 2811 if (tcol) { 2812 *mat_j++ = tcol - 1; 2813 *mat_a++ = rbuf4_i[ct2]; 2814 ilen++; 2815 } 2816 } 2817 } else { /* allcolumns */ 2818 for (l=0; l<max2; l++,ct2++) { 2819 *mat_j++ = rbuf3_i[ct2]; /* same global column index of C */ 2820 *mat_a++ = rbuf4_i[ct2]; 2821 ilen++; 2822 } 2823 } 2824 imat_ilen[row] = ilen; 2825 } 2826 } 2827 } 2828 } 2829 2830 /* sort the rows */ 2831 { 2832 PetscInt *imat_ilen,*imat_j,*imat_i; 2833 PetscScalar *imat_a; 2834 2835 for (i=0; i<ismax; i++) { 2836 mat = (Mat_SeqAIJ*)submats[i]->data; 2837 imat_j = mat->j; 2838 imat_i = mat->i; 2839 imat_a = mat->a; 2840 imat_ilen = mat->ilen; 2841 2842 if (allcolumns[i]) continue; 2843 jmax = nrow[i]; 2844 for (j=0; j<jmax; j++) { 2845 PetscInt ilen; 2846 2847 mat_i = imat_i[j]; 2848 mat_a = imat_a + mat_i; 2849 mat_j = imat_j + mat_i; 2850 ilen = imat_ilen[j]; 2851 ierr = PetscSortIntWithScalarArray(ilen,mat_j,mat_a);CHKERRQ(ierr); 2852 } 2853 } 2854 } 2855 2856 ierr = PetscFree(r_status4);CHKERRQ(ierr); 2857 ierr = PetscFree(r_waits4);CHKERRQ(ierr); 2858 if (nrqr) {ierr = MPI_Waitall(nrqr,s_waits4,s_status4);CHKERRQ(ierr);} 2859 ierr = PetscFree(s_waits4);CHKERRQ(ierr); 2860 ierr = PetscFree(s_status4);CHKERRQ(ierr); 2861 2862 /* Restore the indices */ 2863 for (i=0; i<ismax; i++) { 2864 ierr = ISRestoreIndices(isrow[i],irow+i);CHKERRQ(ierr); 2865 if (!allcolumns[i]) { 2866 ierr = ISRestoreIndices(iscol[i],icol+i);CHKERRQ(ierr); 2867 } 2868 } 2869 2870 /* Destroy allocated memory */ 2871 ierr = PetscFree4(irow,icol,nrow,ncol);CHKERRQ(ierr); 2872 ierr = PetscFree4(w1,w2,w3,w4);CHKERRQ(ierr); 2873 ierr = PetscFree(pa);CHKERRQ(ierr); 2874 2875 ierr = PetscFree4(sbuf1,ptr,tmp,ctr);CHKERRQ(ierr); 2876 ierr = PetscFree(rbuf2);CHKERRQ(ierr); 2877 for (i=0; i<nrqr; ++i) { 2878 ierr = PetscFree(sbuf2[i]);CHKERRQ(ierr); 2879 } 2880 for (i=0; i<nrqs; ++i) { 2881 ierr = PetscFree(rbuf3[i]);CHKERRQ(ierr); 2882 ierr = PetscFree(rbuf4[i]);CHKERRQ(ierr); 2883 } 2884 2885 ierr = PetscFree3(sbuf2,req_size,req_source);CHKERRQ(ierr); 2886 ierr = PetscFree(rbuf3);CHKERRQ(ierr); 2887 ierr = PetscFree(rbuf4);CHKERRQ(ierr); 2888 ierr = PetscFree(sbuf_aj[0]);CHKERRQ(ierr); 2889 ierr = PetscFree(sbuf_aj);CHKERRQ(ierr); 2890 ierr = PetscFree(sbuf_aa[0]);CHKERRQ(ierr); 2891 ierr = PetscFree(sbuf_aa);CHKERRQ(ierr); 2892 2893 #if defined(PETSC_USE_CTABLE) 2894 for (i=0; i<ismax; i++) {ierr = PetscTableDestroy((PetscTable*)&rmap[i]);CHKERRQ(ierr);} 2895 #else 2896 if (ismax) {ierr = PetscFree(rmap[0]);CHKERRQ(ierr);} 2897 #endif 2898 ierr = PetscFree(rmap);CHKERRQ(ierr); 2899 2900 for (i=0; i<ismax; i++) { 2901 if (!allcolumns[i]) { 2902 #if defined(PETSC_USE_CTABLE) 2903 ierr = PetscTableDestroy((PetscTable*)&cmap[i]);CHKERRQ(ierr); 2904 #else 2905 ierr = PetscFree(cmap[i]);CHKERRQ(ierr); 2906 #endif 2907 } 2908 } 2909 ierr = PetscFree(cmap);CHKERRQ(ierr); 2910 if (ismax) {ierr = PetscFree(lens[0]);CHKERRQ(ierr);} 2911 ierr = PetscFree(lens);CHKERRQ(ierr); 2912 2913 for (i=0; i<ismax; i++) { 2914 ierr = MatAssemblyBegin(submats[i],MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2915 ierr = MatAssemblyEnd(submats[i],MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2916 } 2917 2918 ierr = PetscFree(issorted);CHKERRQ(ierr); 2919 ierr = PetscFree(row2proc);CHKERRQ(ierr); 2920 ierr = PetscFree(smats);CHKERRQ(ierr); 2921 PetscFunctionReturn(0); 2922 } 2923 2924 /* 2925 Permute A & B into C's *local* index space using rowemb,dcolemb for A and rowemb,ocolemb for B. 2926 Embeddings are supposed to be injections and the above implies that the range of rowemb is a subset 2927 of [0,m), dcolemb is in [0,n) and ocolemb is in [N-n). 2928 If pattern == DIFFERENT_NONZERO_PATTERN, C is preallocated according to A&B. 2929 After that B's columns are mapped into C's global column space, so that C is in the "disassembled" 2930 state, and needs to be "assembled" later by compressing B's column space. 2931 2932 This function may be called in lieu of preallocation, so C should not be expected to be preallocated. 2933 Following this call, C->A & C->B have been created, even if empty. 2934 */ 2935 PetscErrorCode MatSetSeqMats_MPIAIJ(Mat C,IS rowemb,IS dcolemb,IS ocolemb,MatStructure pattern,Mat A,Mat B) 2936 { 2937 /* If making this function public, change the error returned in this function away from _PLIB. */ 2938 PetscErrorCode ierr; 2939 Mat_MPIAIJ *aij; 2940 Mat_SeqAIJ *Baij; 2941 PetscBool seqaij,Bdisassembled; 2942 PetscInt m,n,*nz,i,j,ngcol,col,rstart,rend,shift,count; 2943 PetscScalar v; 2944 const PetscInt *rowindices,*colindices; 2945 2946 PetscFunctionBegin; 2947 /* Check to make sure the component matrices (and embeddings) are compatible with C. */ 2948 if (A) { 2949 ierr = PetscObjectTypeCompare((PetscObject)A,MATSEQAIJ,&seqaij);CHKERRQ(ierr); 2950 if (!seqaij) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Diagonal matrix is of wrong type"); 2951 if (rowemb) { 2952 ierr = ISGetLocalSize(rowemb,&m);CHKERRQ(ierr); 2953 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); 2954 } else { 2955 if (C->rmap->n != A->rmap->n) { 2956 SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Diag seq matrix is row-incompatible with the MPIAIJ matrix"); 2957 } 2958 } 2959 if (dcolemb) { 2960 ierr = ISGetLocalSize(dcolemb,&n);CHKERRQ(ierr); 2961 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); 2962 } else { 2963 if (C->cmap->n != A->cmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Diag seq matrix is col-incompatible with the MPIAIJ matrix"); 2964 } 2965 } 2966 if (B) { 2967 ierr = PetscObjectTypeCompare((PetscObject)B,MATSEQAIJ,&seqaij);CHKERRQ(ierr); 2968 if (!seqaij) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Off-diagonal matrix is of wrong type"); 2969 if (rowemb) { 2970 ierr = ISGetLocalSize(rowemb,&m);CHKERRQ(ierr); 2971 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); 2972 } else { 2973 if (C->rmap->n != B->rmap->n) { 2974 SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Off-diag seq matrix is row-incompatible with the MPIAIJ matrix"); 2975 } 2976 } 2977 if (ocolemb) { 2978 ierr = ISGetLocalSize(ocolemb,&n);CHKERRQ(ierr); 2979 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); 2980 } else { 2981 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"); 2982 } 2983 } 2984 2985 aij = (Mat_MPIAIJ*)(C->data); 2986 if (!aij->A) { 2987 /* Mimic parts of MatMPIAIJSetPreallocation() */ 2988 ierr = MatCreate(PETSC_COMM_SELF,&aij->A);CHKERRQ(ierr); 2989 ierr = MatSetSizes(aij->A,C->rmap->n,C->cmap->n,C->rmap->n,C->cmap->n);CHKERRQ(ierr); 2990 ierr = MatSetBlockSizesFromMats(aij->A,C,C);CHKERRQ(ierr); 2991 ierr = MatSetType(aij->A,MATSEQAIJ);CHKERRQ(ierr); 2992 ierr = PetscLogObjectParent((PetscObject)C,(PetscObject)aij->A);CHKERRQ(ierr); 2993 } 2994 if (A) { 2995 ierr = MatSetSeqMat_SeqAIJ(aij->A,rowemb,dcolemb,pattern,A);CHKERRQ(ierr); 2996 } else { 2997 ierr = MatSetUp(aij->A);CHKERRQ(ierr); 2998 } 2999 if (B) { /* Destroy the old matrix or the column map, depending on the sparsity pattern. */ 3000 /* 3001 If pattern == DIFFERENT_NONZERO_PATTERN, we reallocate B and 3002 need to "disassemble" B -- convert it to using C's global indices. 3003 To insert the values we take the safer, albeit more expensive, route of MatSetValues(). 3004 3005 If pattern == SUBSET_NONZERO_PATTERN, we do not "disassemble" B and do not reallocate; 3006 we MatZeroValues(B) first, so there may be a bunch of zeros that, perhaps, could be compacted out. 3007 3008 TODO: Put B's values into aij->B's aij structure in place using the embedding ISs? 3009 At least avoid calling MatSetValues() and the implied searches? 3010 */ 3011 3012 if (B && pattern == DIFFERENT_NONZERO_PATTERN) { 3013 #if defined(PETSC_USE_CTABLE) 3014 ierr = PetscTableDestroy(&aij->colmap);CHKERRQ(ierr); 3015 #else 3016 ierr = PetscFree(aij->colmap);CHKERRQ(ierr); 3017 /* A bit of a HACK: ideally we should deal with case aij->B all in one code block below. */ 3018 if (aij->B) { 3019 ierr = PetscLogObjectMemory((PetscObject)C,-aij->B->cmap->n*sizeof(PetscInt));CHKERRQ(ierr); 3020 } 3021 #endif 3022 ngcol = 0; 3023 if (aij->lvec) { 3024 ierr = VecGetSize(aij->lvec,&ngcol);CHKERRQ(ierr); 3025 } 3026 if (aij->garray) { 3027 ierr = PetscFree(aij->garray);CHKERRQ(ierr); 3028 ierr = PetscLogObjectMemory((PetscObject)C,-ngcol*sizeof(PetscInt));CHKERRQ(ierr); 3029 } 3030 ierr = VecDestroy(&aij->lvec);CHKERRQ(ierr); 3031 ierr = VecScatterDestroy(&aij->Mvctx);CHKERRQ(ierr); 3032 } 3033 if (aij->B && B && pattern == DIFFERENT_NONZERO_PATTERN) { 3034 ierr = MatDestroy(&aij->B);CHKERRQ(ierr); 3035 } 3036 if (aij->B && B && pattern == SUBSET_NONZERO_PATTERN) { 3037 ierr = MatZeroEntries(aij->B);CHKERRQ(ierr); 3038 } 3039 } 3040 Bdisassembled = PETSC_FALSE; 3041 if (!aij->B) { 3042 ierr = MatCreate(PETSC_COMM_SELF,&aij->B);CHKERRQ(ierr); 3043 ierr = PetscLogObjectParent((PetscObject)C,(PetscObject)aij->B);CHKERRQ(ierr); 3044 ierr = MatSetSizes(aij->B,C->rmap->n,C->cmap->N,C->rmap->n,C->cmap->N);CHKERRQ(ierr); 3045 ierr = MatSetBlockSizesFromMats(aij->B,B,B);CHKERRQ(ierr); 3046 ierr = MatSetType(aij->B,MATSEQAIJ);CHKERRQ(ierr); 3047 Bdisassembled = PETSC_TRUE; 3048 } 3049 if (B) { 3050 Baij = (Mat_SeqAIJ*)(B->data); 3051 if (pattern == DIFFERENT_NONZERO_PATTERN) { 3052 ierr = PetscMalloc1(B->rmap->n,&nz);CHKERRQ(ierr); 3053 for (i=0; i<B->rmap->n; i++) { 3054 nz[i] = Baij->i[i+1] - Baij->i[i]; 3055 } 3056 ierr = MatSeqAIJSetPreallocation(aij->B,0,nz);CHKERRQ(ierr); 3057 ierr = PetscFree(nz);CHKERRQ(ierr); 3058 } 3059 3060 ierr = PetscLayoutGetRange(C->rmap,&rstart,&rend);CHKERRQ(ierr); 3061 shift = rend-rstart; 3062 count = 0; 3063 rowindices = NULL; 3064 colindices = NULL; 3065 if (rowemb) { 3066 ierr = ISGetIndices(rowemb,&rowindices);CHKERRQ(ierr); 3067 } 3068 if (ocolemb) { 3069 ierr = ISGetIndices(ocolemb,&colindices);CHKERRQ(ierr); 3070 } 3071 for (i=0; i<B->rmap->n; i++) { 3072 PetscInt row; 3073 row = i; 3074 if (rowindices) row = rowindices[i]; 3075 for (j=Baij->i[i]; j<Baij->i[i+1]; j++) { 3076 col = Baij->j[count]; 3077 if (colindices) col = colindices[col]; 3078 if (Bdisassembled && col>=rstart) col += shift; 3079 v = Baij->a[count]; 3080 ierr = MatSetValues(aij->B,1,&row,1,&col,&v,INSERT_VALUES);CHKERRQ(ierr); 3081 ++count; 3082 } 3083 } 3084 /* No assembly for aij->B is necessary. */ 3085 /* FIXME: set aij->B's nonzerostate correctly. */ 3086 } else { 3087 ierr = MatSetUp(aij->B);CHKERRQ(ierr); 3088 } 3089 C->preallocated = PETSC_TRUE; 3090 C->was_assembled = PETSC_FALSE; 3091 C->assembled = PETSC_FALSE; 3092 /* 3093 C will need to be assembled so that aij->B can be compressed into local form in MatSetUpMultiply_MPIAIJ(). 3094 Furthermore, its nonzerostate will need to be based on that of aij->A's and aij->B's. 3095 */ 3096 PetscFunctionReturn(0); 3097 } 3098 3099 /* 3100 B uses local indices with column indices ranging between 0 and N-n; they must be interpreted using garray. 3101 */ 3102 PetscErrorCode MatGetSeqMats_MPIAIJ(Mat C,Mat *A,Mat *B) 3103 { 3104 Mat_MPIAIJ *aij = (Mat_MPIAIJ*) (C->data); 3105 3106 PetscFunctionBegin; 3107 PetscValidPointer(A,2); 3108 PetscValidPointer(B,3); 3109 /* FIXME: make sure C is assembled */ 3110 *A = aij->A; 3111 *B = aij->B; 3112 /* Note that we don't incref *A and *B, so be careful! */ 3113 PetscFunctionReturn(0); 3114 } 3115 3116 /* 3117 Extract MPI submatrices encoded by pairs of IS that may live on subcomms of C. 3118 NOT SCALABLE due to the use of ISGetNonlocalIS() (see below). 3119 */ 3120 PetscErrorCode MatGetSubMatricesMPI_MPIXAIJ(Mat C,PetscInt ismax,const IS isrow[],const IS iscol[],MatReuse scall,Mat *submat[], 3121 PetscErrorCode(*getsubmats_seq)(Mat,PetscInt,const IS[],const IS[],MatReuse,Mat**), 3122 PetscErrorCode(*getlocalmats)(Mat,Mat*,Mat*), 3123 PetscErrorCode(*setseqmat)(Mat,IS,IS,MatStructure,Mat), 3124 PetscErrorCode(*setseqmats)(Mat,IS,IS,IS,MatStructure,Mat,Mat)) 3125 { 3126 PetscErrorCode ierr; 3127 PetscMPIInt isize,flag; 3128 PetscInt i,ii,cismax,ispar; 3129 Mat *A,*B; 3130 IS *isrow_p,*iscol_p,*cisrow,*ciscol,*ciscol_p; 3131 3132 PetscFunctionBegin; 3133 if (!ismax) PetscFunctionReturn(0); 3134 3135 for (i = 0, cismax = 0; i < ismax; ++i) { 3136 PetscMPIInt isize; 3137 ierr = MPI_Comm_compare(((PetscObject)isrow[i])->comm,((PetscObject)iscol[i])->comm,&flag);CHKERRQ(ierr); 3138 if (flag != MPI_IDENT) SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Row and column index sets must have the same communicator"); 3139 ierr = MPI_Comm_size(((PetscObject)isrow[i])->comm, &isize);CHKERRQ(ierr); 3140 if (isize > 1) ++cismax; 3141 } 3142 3143 /* 3144 If cismax is zero on all C's ranks, then and only then can we use purely sequential matrix extraction. 3145 ispar counts the number of parallel ISs across C's comm. 3146 */ 3147 ierr = MPIU_Allreduce(&cismax,&ispar,1,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)C));CHKERRQ(ierr); 3148 if (!ispar) { /* Sequential ISs only across C's comm, so can call the sequential matrix extraction subroutine. */ 3149 ierr = (*getsubmats_seq)(C,ismax,isrow,iscol,scall,submat);CHKERRQ(ierr); 3150 PetscFunctionReturn(0); 3151 } 3152 3153 /* if (ispar) */ 3154 /* 3155 Construct the "complements" -- the off-processor indices -- of the iscol ISs for parallel ISs only. 3156 These are used to extract the off-diag portion of the resulting parallel matrix. 3157 The row IS for the off-diag portion is the same as for the diag portion, 3158 so we merely alias (without increfing) the row IS, while skipping those that are sequential. 3159 */ 3160 ierr = PetscMalloc2(cismax,&cisrow,cismax,&ciscol);CHKERRQ(ierr); 3161 ierr = PetscMalloc1(cismax,&ciscol_p);CHKERRQ(ierr); 3162 for (i = 0, ii = 0; i < ismax; ++i) { 3163 ierr = MPI_Comm_size(((PetscObject)isrow[i])->comm,&isize);CHKERRQ(ierr); 3164 if (isize > 1) { 3165 /* 3166 TODO: This is the part that's ***NOT SCALABLE***. 3167 To fix this we need to extract just the indices of C's nonzero columns 3168 that lie on the intersection of isrow[i] and ciscol[ii] -- the nonlocal 3169 part of iscol[i] -- without actually computing ciscol[ii]. This also has 3170 to be done without serializing on the IS list, so, most likely, it is best 3171 done by rewriting MatGetSubMatrices_MPIAIJ() directly. 3172 */ 3173 ierr = ISGetNonlocalIS(iscol[i],&(ciscol[ii]));CHKERRQ(ierr); 3174 /* Now we have to 3175 (a) make sure ciscol[ii] is sorted, since, even if the off-proc indices 3176 were sorted on each rank, concatenated they might no longer be sorted; 3177 (b) Use ISSortPermutation() to construct ciscol_p, the mapping from the 3178 indices in the nondecreasing order to the original index positions. 3179 If ciscol[ii] is strictly increasing, the permutation IS is NULL. 3180 */ 3181 ierr = ISSortPermutation(ciscol[ii],PETSC_FALSE,ciscol_p+ii);CHKERRQ(ierr); 3182 ierr = ISSort(ciscol[ii]);CHKERRQ(ierr); 3183 ++ii; 3184 } 3185 } 3186 ierr = PetscMalloc2(ismax,&isrow_p,ismax,&iscol_p);CHKERRQ(ierr); 3187 for (i = 0, ii = 0; i < ismax; ++i) { 3188 PetscInt j,issize; 3189 const PetscInt *indices; 3190 3191 /* 3192 Permute the indices into a nondecreasing order. Reject row and col indices with duplicates. 3193 */ 3194 ierr = ISSortPermutation(isrow[i],PETSC_FALSE,isrow_p+i);CHKERRQ(ierr); 3195 ierr = ISSort(isrow[i]);CHKERRQ(ierr); 3196 ierr = ISGetLocalSize(isrow[i],&issize);CHKERRQ(ierr); 3197 ierr = ISGetIndices(isrow[i],&indices);CHKERRQ(ierr); 3198 for (j = 1; j < issize; ++j) { 3199 if (indices[j] == indices[j-1]) { 3200 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]); 3201 } 3202 } 3203 ierr = ISRestoreIndices(isrow[i],&indices);CHKERRQ(ierr); 3204 3205 3206 ierr = ISSortPermutation(iscol[i],PETSC_FALSE,iscol_p+i);CHKERRQ(ierr); 3207 ierr = ISSort(iscol[i]);CHKERRQ(ierr); 3208 ierr = ISGetLocalSize(iscol[i],&issize);CHKERRQ(ierr); 3209 ierr = ISGetIndices(iscol[i],&indices);CHKERRQ(ierr); 3210 for (j = 1; j < issize; ++j) { 3211 if (indices[j-1] == indices[j]) { 3212 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]); 3213 } 3214 } 3215 ierr = ISRestoreIndices(iscol[i],&indices);CHKERRQ(ierr); 3216 ierr = MPI_Comm_size(((PetscObject)isrow[i])->comm,&isize);CHKERRQ(ierr); 3217 if (isize > 1) { 3218 cisrow[ii] = isrow[i]; 3219 ++ii; 3220 } 3221 } 3222 /* 3223 Allocate the necessary arrays to hold the resulting parallel matrices as well as the intermediate 3224 array of sequential matrices underlying the resulting parallel matrices. 3225 Which arrays to allocate is based on the value of MatReuse scall and whether ISs are sorted and/or 3226 contain duplicates. 3227 3228 There are as many diag matrices as there are original index sets. There are only as many parallel 3229 and off-diag matrices, as there are parallel (comm size > 1) index sets. 3230 3231 ARRAYS that can hold Seq matrices get allocated in any event -- either here or by getsubmats_seq(): 3232 - If the array of MPI matrices already exists and is being reused, we need to allocate the array 3233 and extract the underlying seq matrices into it to serve as placeholders, into which getsubmats_seq 3234 will deposite the extracted diag and off-diag parts. Thus, we allocate the A&B arrays and fill them 3235 with A[i] and B[ii] extracted from the corresponding MPI submat. 3236 - However, if the rows, A's column indices or B's column indices are not sorted, the extracted A[i] & B[ii] 3237 will have a different order from what getsubmats_seq expects. To handle this case -- indicated 3238 by a nonzero isrow_p[i], iscol_p[i], or ciscol_p[ii] -- we duplicate A[i] --> AA[i], B[ii] --> BB[ii] 3239 (retrieve composed AA[i] or BB[ii]) and reuse them here. AA[i] and BB[ii] are then used to permute its 3240 values into A[i] and B[ii] sitting inside the corresponding submat. 3241 - If no reuse is taking place then getsubmats_seq will allocate the A&B arrays and create the corresponding 3242 A[i], B[ii], AA[i] or BB[ii] matrices. 3243 */ 3244 /* Parallel matrix array is allocated here only if no reuse is taking place. If reused, it is passed in by the caller. */ 3245 if (scall == MAT_INITIAL_MATRIX) { 3246 ierr = PetscMalloc1(ismax,submat);CHKERRQ(ierr); 3247 } 3248 3249 /* Now obtain the sequential A and B submatrices separately. */ 3250 /* scall=MAT_REUSE_MATRIX is not handled yet, because getsubmats_seq() requires reuse of A and B */ 3251 ierr = (*getsubmats_seq)(C,ismax,isrow,iscol,MAT_INITIAL_MATRIX,&A);CHKERRQ(ierr); 3252 ierr = (*getsubmats_seq)(C,cismax,cisrow,ciscol,MAT_INITIAL_MATRIX,&B);CHKERRQ(ierr); 3253 3254 /* 3255 If scall == MAT_REUSE_MATRIX AND the permutations are NULL, we are done, since the sequential 3256 matrices A & B have been extracted directly into the parallel matrices containing them, or 3257 simply into the sequential matrix identical with the corresponding A (if isize == 1). 3258 Note that in that case colmap doesn't need to be rebuilt, since the matrices are expected 3259 to have the same sparsity pattern. 3260 Otherwise, A and/or B have to be properly embedded into C's index spaces and the correct colmap 3261 must be constructed for C. This is done by setseqmat(s). 3262 */ 3263 for (i = 0, ii = 0; i < ismax; ++i) { 3264 /* 3265 TODO: cache ciscol, permutation ISs and maybe cisrow? What about isrow & iscol? 3266 That way we can avoid sorting and computing permutations when reusing. 3267 To this end: 3268 - remove the old cache, if it exists, when extracting submatrices with MAT_INITIAL_MATRIX 3269 - if caching arrays to hold the ISs, make and compose a container for them so that it can 3270 be destroyed upon destruction of C (use PetscContainerUserDestroy() to clear out the contents). 3271 */ 3272 MatStructure pattern; 3273 pattern = DIFFERENT_NONZERO_PATTERN; 3274 3275 ierr = MPI_Comm_size(((PetscObject)isrow[i])->comm,&isize);CHKERRQ(ierr); 3276 /* Construct submat[i] from the Seq pieces A (and B, if necessary). */ 3277 if (isize > 1) { 3278 if (scall == MAT_INITIAL_MATRIX) { 3279 ierr = MatCreate(((PetscObject)isrow[i])->comm,(*submat)+i);CHKERRQ(ierr); 3280 ierr = MatSetSizes((*submat)[i],A[i]->rmap->n,A[i]->cmap->n,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 3281 ierr = MatSetType((*submat)[i],MATMPIAIJ);CHKERRQ(ierr); 3282 ierr = PetscLayoutSetUp((*submat)[i]->rmap);CHKERRQ(ierr); 3283 ierr = PetscLayoutSetUp((*submat)[i]->cmap);CHKERRQ(ierr); 3284 } 3285 /* 3286 For each parallel isrow[i], insert the extracted sequential matrices into the parallel matrix. 3287 */ 3288 { 3289 Mat AA,BB; 3290 AA = A[i]; 3291 BB = B[ii]; 3292 if (AA || BB) { 3293 ierr = setseqmats((*submat)[i],isrow_p[i],iscol_p[i],ciscol_p[ii],pattern,AA,BB);CHKERRQ(ierr); 3294 ierr = MatAssemblyBegin((*submat)[i],MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3295 ierr = MatAssemblyEnd((*submat)[i],MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3296 } 3297 3298 ierr = MatDestroy(&AA);CHKERRQ(ierr); 3299 ierr = MatDestroy(&BB);CHKERRQ(ierr); 3300 } 3301 ierr = ISDestroy(ciscol+ii);CHKERRQ(ierr); 3302 ierr = ISDestroy(ciscol_p+ii);CHKERRQ(ierr); 3303 ++ii; 3304 } else { /* if (isize == 1) */ 3305 if (scall == MAT_REUSE_MATRIX) { 3306 ierr = MatDestroy(&(*submat)[i]);CHKERRQ(ierr); 3307 } 3308 if (isrow_p[i] || iscol_p[i]) { 3309 ierr = MatDuplicate(A[i],MAT_DO_NOT_COPY_VALUES,(*submat)+i);CHKERRQ(ierr); 3310 ierr = setseqmat((*submat)[i],isrow_p[i],iscol_p[i],pattern,A[i]);CHKERRQ(ierr); 3311 /* Otherwise A is extracted straight into (*submats)[i]. */ 3312 /* TODO: Compose A[i] on (*submat([i] for future use, if ((isrow_p[i] || iscol_p[i]) && MAT_INITIAL_MATRIX). */ 3313 ierr = MatDestroy(A+i);CHKERRQ(ierr); 3314 } else (*submat)[i] = A[i]; 3315 } 3316 ierr = ISDestroy(&isrow_p[i]);CHKERRQ(ierr); 3317 ierr = ISDestroy(&iscol_p[i]);CHKERRQ(ierr); 3318 } 3319 ierr = PetscFree2(cisrow,ciscol);CHKERRQ(ierr); 3320 ierr = PetscFree2(isrow_p,iscol_p);CHKERRQ(ierr); 3321 ierr = PetscFree(ciscol_p);CHKERRQ(ierr); 3322 ierr = PetscFree(A);CHKERRQ(ierr); 3323 ierr = PetscFree(B);CHKERRQ(ierr); 3324 PetscFunctionReturn(0); 3325 } 3326 3327 PetscErrorCode MatGetSubMatricesMPI_MPIAIJ(Mat C,PetscInt ismax,const IS isrow[],const IS iscol[],MatReuse scall,Mat *submat[]) 3328 { 3329 PetscErrorCode ierr; 3330 3331 PetscFunctionBegin; 3332 ierr = MatGetSubMatricesMPI_MPIXAIJ(C,ismax,isrow,iscol,scall,submat,MatGetSubMatrices_MPIAIJ,MatGetSeqMats_MPIAIJ,MatSetSeqMat_SeqAIJ,MatSetSeqMats_MPIAIJ);CHKERRQ(ierr); 3333 PetscFunctionReturn(0); 3334 } 3335