1 #define PETSCMAT_DLL 2 3 /* 4 Support for the parallel AIJ matrix vector multiply 5 */ 6 #include "src/mat/impls/aij/mpi/mpiaij.h" 7 8 #undef __FUNCT__ 9 #define __FUNCT__ "MatSetUpMultiply_MPIAIJ" 10 PetscErrorCode MatSetUpMultiply_MPIAIJ(Mat mat) 11 { 12 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 13 Mat_SeqAIJ *B = (Mat_SeqAIJ*)(aij->B->data); 14 PetscErrorCode ierr; 15 PetscInt i,j,*aj = B->j,ec = 0,*garray; 16 IS from,to; 17 Vec gvec; 18 PetscTruth useblockis; 19 #if defined (PETSC_USE_CTABLE) 20 PetscTable gid1_lid1; 21 PetscTablePosition tpos; 22 PetscInt gid,lid; 23 #else 24 PetscInt N = mat->N,*indices; 25 #endif 26 27 PetscFunctionBegin; 28 29 #if defined (PETSC_USE_CTABLE) 30 /* use a table - Mark Adams */ 31 ierr = PetscTableCreate(aij->B->m,&gid1_lid1);CHKERRQ(ierr); 32 for (i=0; i<aij->B->m; i++) { 33 for (j=0; j<B->ilen[i]; j++) { 34 PetscInt data,gid1 = aj[B->i[i] + j] + 1; 35 ierr = PetscTableFind(gid1_lid1,gid1,&data);CHKERRQ(ierr); 36 if (!data) { 37 /* one based table */ 38 ierr = PetscTableAdd(gid1_lid1,gid1,++ec);CHKERRQ(ierr); 39 } 40 } 41 } 42 /* form array of columns we need */ 43 ierr = PetscMalloc((ec+1)*sizeof(PetscInt),&garray);CHKERRQ(ierr); 44 ierr = PetscTableGetHeadPosition(gid1_lid1,&tpos);CHKERRQ(ierr); 45 while (tpos) { 46 ierr = PetscTableGetNext(gid1_lid1,&tpos,&gid,&lid);CHKERRQ(ierr); 47 gid--; 48 lid--; 49 garray[lid] = gid; 50 } 51 ierr = PetscSortInt(ec,garray);CHKERRQ(ierr); /* sort, and rebuild */ 52 ierr = PetscTableRemoveAll(gid1_lid1);CHKERRQ(ierr); 53 for (i=0; i<ec; i++) { 54 ierr = PetscTableAdd(gid1_lid1,garray[i]+1,i+1);CHKERRQ(ierr); 55 } 56 /* compact out the extra columns in B */ 57 for (i=0; i<aij->B->m; i++) { 58 for (j=0; j<B->ilen[i]; j++) { 59 PetscInt gid1 = aj[B->i[i] + j] + 1; 60 ierr = PetscTableFind(gid1_lid1,gid1,&lid);CHKERRQ(ierr); 61 lid --; 62 aj[B->i[i] + j] = lid; 63 } 64 } 65 aij->B->n = aij->B->N = ec; 66 ierr = PetscTableDelete(gid1_lid1);CHKERRQ(ierr); 67 /* Mark Adams */ 68 #else 69 /* For the first stab we make an array as long as the number of columns */ 70 /* mark those columns that are in aij->B */ 71 ierr = PetscMalloc((N+1)*sizeof(PetscInt),&indices);CHKERRQ(ierr); 72 ierr = PetscMemzero(indices,N*sizeof(PetscInt));CHKERRQ(ierr); 73 for (i=0; i<aij->B->m; i++) { 74 for (j=0; j<B->ilen[i]; j++) { 75 if (!indices[aj[B->i[i] + j] ]) ec++; 76 indices[aj[B->i[i] + j] ] = 1; 77 } 78 } 79 80 /* form array of columns we need */ 81 ierr = PetscMalloc((ec+1)*sizeof(PetscInt),&garray);CHKERRQ(ierr); 82 ec = 0; 83 for (i=0; i<N; i++) { 84 if (indices[i]) garray[ec++] = i; 85 } 86 87 /* make indices now point into garray */ 88 for (i=0; i<ec; i++) { 89 indices[garray[i]] = i; 90 } 91 92 /* compact out the extra columns in B */ 93 for (i=0; i<aij->B->m; i++) { 94 for (j=0; j<B->ilen[i]; j++) { 95 aj[B->i[i] + j] = indices[aj[B->i[i] + j]]; 96 } 97 } 98 aij->B->n = aij->B->N = ec; 99 ierr = PetscFree(indices);CHKERRQ(ierr); 100 #endif 101 /* create local vector that is used to scatter into */ 102 ierr = VecCreateSeq(PETSC_COMM_SELF,ec,&aij->lvec);CHKERRQ(ierr); 103 104 /* create two temporary Index sets for build scatter gather */ 105 /* check for the special case where blocks are communicated for faster VecScatterXXX */ 106 useblockis = PETSC_TRUE; 107 if (mat->bs > 1) { 108 PetscInt bs = mat->bs,ibs,ga; 109 if (!(ec % bs)) { 110 for (i=0; i<ec/bs; i++) { 111 if ((ga = garray[ibs = i*bs]) % bs) { 112 useblockis = PETSC_FALSE; 113 break; 114 } 115 for (j=1; j<bs; j++) { 116 if (garray[ibs+j] != ga+j) { 117 useblockis = PETSC_FALSE; 118 break; 119 } 120 } 121 if (!useblockis) break; 122 } 123 } 124 } 125 if (useblockis) { 126 PetscInt *ga,bs = mat->bs,iec = ec/bs; 127 ierr = PetscLogInfo((mat,"MatSetUpMultiply_MPIAIJ: Using block index set to define scatter\n")); 128 ierr = PetscMalloc((ec/mat->bs)*sizeof(PetscInt),&ga);CHKERRQ(ierr); 129 for (i=0; i<iec; i++) ga[i] = garray[i*bs]; 130 ierr = ISCreateBlock(mat->comm,bs,iec,ga,&from);CHKERRQ(ierr); 131 ierr = PetscFree(ga);CHKERRQ(ierr); 132 } else { 133 ierr = ISCreateGeneral(mat->comm,ec,garray,&from);CHKERRQ(ierr); 134 } 135 ierr = ISCreateStride(PETSC_COMM_SELF,ec,0,1,&to);CHKERRQ(ierr); 136 137 /* create temporary global vector to generate scatter context */ 138 /* this is inefficient, but otherwise we must do either 139 1) save garray until the first actual scatter when the vector is known or 140 2) have another way of generating a scatter context without a vector.*/ 141 ierr = VecCreateMPI(mat->comm,mat->n,mat->N,&gvec);CHKERRQ(ierr); 142 143 /* generate the scatter context */ 144 ierr = VecScatterCreate(gvec,from,aij->lvec,to,&aij->Mvctx);CHKERRQ(ierr); 145 ierr = PetscLogObjectParent(mat,aij->Mvctx);CHKERRQ(ierr); 146 ierr = PetscLogObjectParent(mat,aij->lvec);CHKERRQ(ierr); 147 ierr = PetscLogObjectParent(mat,from);CHKERRQ(ierr); 148 ierr = PetscLogObjectParent(mat,to);CHKERRQ(ierr); 149 aij->garray = garray; 150 ierr = PetscLogObjectMemory(mat,(ec+1)*sizeof(PetscInt));CHKERRQ(ierr); 151 ierr = ISDestroy(from);CHKERRQ(ierr); 152 ierr = ISDestroy(to);CHKERRQ(ierr); 153 ierr = VecDestroy(gvec);CHKERRQ(ierr); 154 PetscFunctionReturn(0); 155 } 156 157 158 #undef __FUNCT__ 159 #define __FUNCT__ "DisAssemble_MPIAIJ" 160 /* 161 Takes the local part of an already assembled MPIAIJ matrix 162 and disassembles it. This is to allow new nonzeros into the matrix 163 that require more communication in the matrix vector multiply. 164 Thus certain data-structures must be rebuilt. 165 166 Kind of slow! But that's what application programmers get when 167 they are sloppy. 168 */ 169 PetscErrorCode DisAssemble_MPIAIJ(Mat A) 170 { 171 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)A->data; 172 Mat B = aij->B,Bnew; 173 Mat_SeqAIJ *Baij = (Mat_SeqAIJ*)B->data; 174 PetscErrorCode ierr; 175 PetscInt i,j,m = B->m,n = A->N,col,ct = 0,*garray = aij->garray,*nz,ec; 176 PetscScalar v; 177 178 PetscFunctionBegin; 179 /* free stuff related to matrix-vec multiply */ 180 ierr = VecGetSize(aij->lvec,&ec);CHKERRQ(ierr); /* needed for PetscLogObjectMemory below */ 181 ierr = VecDestroy(aij->lvec);CHKERRQ(ierr); aij->lvec = 0; 182 ierr = VecScatterDestroy(aij->Mvctx);CHKERRQ(ierr); aij->Mvctx = 0; 183 if (aij->colmap) { 184 #if defined (PETSC_USE_CTABLE) 185 ierr = PetscTableDelete(aij->colmap);CHKERRQ(ierr); 186 aij->colmap = 0; 187 #else 188 ierr = PetscFree(aij->colmap);CHKERRQ(ierr); 189 aij->colmap = 0; 190 ierr = PetscLogObjectMemory(A,-aij->B->n*sizeof(PetscInt));CHKERRQ(ierr); 191 #endif 192 } 193 194 /* make sure that B is assembled so we can access its values */ 195 ierr = MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 196 ierr = MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 197 198 /* invent new B and copy stuff over */ 199 ierr = PetscMalloc((m+1)*sizeof(PetscInt),&nz);CHKERRQ(ierr); 200 for (i=0; i<m; i++) { 201 nz[i] = Baij->i[i+1] - Baij->i[i]; 202 } 203 ierr = MatCreate(PETSC_COMM_SELF,m,n,m,n,&Bnew);CHKERRQ(ierr); 204 ierr = MatSetType(Bnew,B->type_name);CHKERRQ(ierr); 205 ierr = MatSeqAIJSetPreallocation(Bnew,0,nz);CHKERRQ(ierr); 206 ierr = PetscFree(nz);CHKERRQ(ierr); 207 for (i=0; i<m; i++) { 208 for (j=Baij->i[i]; j<Baij->i[i+1]; j++) { 209 col = garray[Baij->j[ct]]; 210 v = Baij->a[ct++]; 211 ierr = MatSetValues(Bnew,1,&i,1,&col,&v,B->insertmode);CHKERRQ(ierr); 212 } 213 } 214 ierr = PetscFree(aij->garray);CHKERRQ(ierr); 215 aij->garray = 0; 216 ierr = PetscLogObjectMemory(A,-ec*sizeof(PetscInt));CHKERRQ(ierr); 217 ierr = MatDestroy(B);CHKERRQ(ierr); 218 ierr = PetscLogObjectParent(A,Bnew);CHKERRQ(ierr); 219 aij->B = Bnew; 220 A->was_assembled = PETSC_FALSE; 221 PetscFunctionReturn(0); 222 } 223 224 /* ugly stuff added for Glenn someday we should fix this up */ 225 226 static PetscInt *auglyrmapd = 0,*auglyrmapo = 0; /* mapping from the local ordering to the "diagonal" and "off-diagonal" 227 parts of the local matrix */ 228 static Vec auglydd = 0,auglyoo = 0; /* work vectors used to scale the two parts of the local matrix */ 229 230 231 #undef __FUNCT__ 232 #define __FUNCT__ "MatMPIAIJDiagonalScaleLocalSetUp" 233 PetscErrorCode MatMPIAIJDiagonalScaleLocalSetUp(Mat inA,Vec scale) 234 { 235 Mat_MPIAIJ *ina = (Mat_MPIAIJ*) inA->data; /*access private part of matrix */ 236 PetscErrorCode ierr; 237 PetscInt i,n,nt,cstart,cend,no,*garray = ina->garray,*lindices; 238 PetscInt *r_rmapd,*r_rmapo; 239 240 PetscFunctionBegin; 241 ierr = MatGetOwnershipRange(inA,&cstart,&cend);CHKERRQ(ierr); 242 ierr = MatGetSize(ina->A,PETSC_NULL,&n);CHKERRQ(ierr); 243 ierr = PetscMalloc((inA->mapping->n+1)*sizeof(PetscInt),&r_rmapd);CHKERRQ(ierr); 244 ierr = PetscMemzero(r_rmapd,inA->mapping->n*sizeof(PetscInt));CHKERRQ(ierr); 245 nt = 0; 246 for (i=0; i<inA->mapping->n; i++) { 247 if (inA->mapping->indices[i] >= cstart && inA->mapping->indices[i] < cend) { 248 nt++; 249 r_rmapd[i] = inA->mapping->indices[i] + 1; 250 } 251 } 252 if (nt != n) SETERRQ2(PETSC_ERR_PLIB,"Hmm nt %D n %D",nt,n); 253 ierr = PetscMalloc((n+1)*sizeof(PetscInt),&auglyrmapd);CHKERRQ(ierr); 254 for (i=0; i<inA->mapping->n; i++) { 255 if (r_rmapd[i]){ 256 auglyrmapd[(r_rmapd[i]-1)-cstart] = i; 257 } 258 } 259 ierr = PetscFree(r_rmapd);CHKERRQ(ierr); 260 ierr = VecCreateSeq(PETSC_COMM_SELF,n,&auglydd);CHKERRQ(ierr); 261 262 ierr = PetscMalloc((inA->N+1)*sizeof(PetscInt),&lindices);CHKERRQ(ierr); 263 ierr = PetscMemzero(lindices,inA->N*sizeof(PetscInt));CHKERRQ(ierr); 264 for (i=0; i<ina->B->n; i++) { 265 lindices[garray[i]] = i+1; 266 } 267 no = inA->mapping->n - nt; 268 ierr = PetscMalloc((inA->mapping->n+1)*sizeof(PetscInt),&r_rmapo);CHKERRQ(ierr); 269 ierr = PetscMemzero(r_rmapo,inA->mapping->n*sizeof(PetscInt));CHKERRQ(ierr); 270 nt = 0; 271 for (i=0; i<inA->mapping->n; i++) { 272 if (lindices[inA->mapping->indices[i]]) { 273 nt++; 274 r_rmapo[i] = lindices[inA->mapping->indices[i]]; 275 } 276 } 277 if (nt > no) SETERRQ2(PETSC_ERR_PLIB,"Hmm nt %D no %D",nt,n); 278 ierr = PetscFree(lindices);CHKERRQ(ierr); 279 ierr = PetscMalloc((nt+1)*sizeof(PetscInt),&auglyrmapo);CHKERRQ(ierr); 280 for (i=0; i<inA->mapping->n; i++) { 281 if (r_rmapo[i]){ 282 auglyrmapo[(r_rmapo[i]-1)] = i; 283 } 284 } 285 ierr = PetscFree(r_rmapo);CHKERRQ(ierr); 286 ierr = VecCreateSeq(PETSC_COMM_SELF,nt,&auglyoo);CHKERRQ(ierr); 287 288 PetscFunctionReturn(0); 289 } 290 291 #undef __FUNCT__ 292 #define __FUNCT__ "MatMPIAIJDiagonalScaleLocal" 293 PetscErrorCode MatMPIAIJDiagonalScaleLocal(Mat A,Vec scale) 294 { 295 /* This routine should really be abandoned as it duplicates MatDiagonalScaleLocal */ 296 PetscErrorCode ierr,(*f)(Mat,Vec); 297 298 PetscFunctionBegin; 299 ierr = PetscObjectQueryFunction((PetscObject)A,"MatDiagonalScaleLocal_C",(void (**)(void))&f);CHKERRQ(ierr); 300 if (f) { 301 ierr = (*f)(A,scale);CHKERRQ(ierr); 302 } 303 PetscFunctionReturn(0); 304 } 305 306 EXTERN_C_BEGIN 307 #undef __FUNCT__ 308 #define __FUNCT__ "MatDiagonalScaleLocal_MPIAIJ" 309 PetscErrorCode PETSCMAT_DLLEXPORT MatDiagonalScaleLocal_MPIAIJ(Mat A,Vec scale) 310 { 311 Mat_MPIAIJ *a = (Mat_MPIAIJ*) A->data; /*access private part of matrix */ 312 PetscErrorCode ierr; 313 PetscInt n,i; 314 PetscScalar *d,*o,*s; 315 316 PetscFunctionBegin; 317 if (!auglyrmapd) { 318 ierr = MatMPIAIJDiagonalScaleLocalSetUp(A,scale);CHKERRQ(ierr); 319 } 320 321 ierr = VecGetArray(scale,&s);CHKERRQ(ierr); 322 323 ierr = VecGetLocalSize(auglydd,&n);CHKERRQ(ierr); 324 ierr = VecGetArray(auglydd,&d);CHKERRQ(ierr); 325 for (i=0; i<n; i++) { 326 d[i] = s[auglyrmapd[i]]; /* copy "diagonal" (true local) portion of scale into dd vector */ 327 } 328 ierr = VecRestoreArray(auglydd,&d);CHKERRQ(ierr); 329 /* column scale "diagonal" portion of local matrix */ 330 ierr = MatDiagonalScale(a->A,PETSC_NULL,auglydd);CHKERRQ(ierr); 331 332 ierr = VecGetLocalSize(auglyoo,&n);CHKERRQ(ierr); 333 ierr = VecGetArray(auglyoo,&o);CHKERRQ(ierr); 334 for (i=0; i<n; i++) { 335 o[i] = s[auglyrmapo[i]]; /* copy "off-diagonal" portion of scale into oo vector */ 336 } 337 ierr = VecRestoreArray(scale,&s);CHKERRQ(ierr); 338 ierr = VecRestoreArray(auglyoo,&o);CHKERRQ(ierr); 339 /* column scale "off-diagonal" portion of local matrix */ 340 ierr = MatDiagonalScale(a->B,PETSC_NULL,auglyoo);CHKERRQ(ierr); 341 342 PetscFunctionReturn(0); 343 } 344 EXTERN_C_END 345 346 347