1 2 #include <petsc-private/matimpl.h> /*I "petscmat.h" I*/ 3 4 #undef __FUNCT__ 5 #define __FUNCT__ "MatAXPY" 6 /*@ 7 MatAXPY - Computes Y = a*X + Y. 8 9 Logically Collective on Mat 10 11 Input Parameters: 12 + a - the scalar multiplier 13 . X - the first matrix 14 . Y - the second matrix 15 - str - either SAME_NONZERO_PATTERN, DIFFERENT_NONZERO_PATTERN 16 or SUBSET_NONZERO_PATTERN (nonzeros of X is a subset of Y's) 17 18 Level: intermediate 19 20 .keywords: matrix, add 21 22 .seealso: MatAYPX() 23 @*/ 24 PetscErrorCode MatAXPY(Mat Y,PetscScalar a,Mat X,MatStructure str) 25 { 26 PetscErrorCode ierr; 27 PetscInt m1,m2,n1,n2; 28 29 PetscFunctionBegin; 30 PetscValidHeaderSpecific(X,MAT_CLASSID,3); 31 PetscValidHeaderSpecific(Y,MAT_CLASSID,1); 32 PetscValidLogicalCollectiveScalar(Y,a,2); 33 ierr = MatGetSize(X,&m1,&n1);CHKERRQ(ierr); 34 ierr = MatGetSize(Y,&m2,&n2);CHKERRQ(ierr); 35 if (m1 != m2 || n1 != n2) SETERRQ4(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Non conforming matrix add: %D %D %D %D",m1,m2,n1,n2); 36 37 ierr = PetscLogEventBegin(MAT_AXPY,Y,0,0,0);CHKERRQ(ierr); 38 if (Y->ops->axpy) { 39 ierr = (*Y->ops->axpy)(Y,a,X,str);CHKERRQ(ierr); 40 } else { 41 ierr = MatAXPY_Basic(Y,a,X,str);CHKERRQ(ierr); 42 } 43 ierr = PetscLogEventEnd(MAT_AXPY,Y,0,0,0);CHKERRQ(ierr); 44 #if defined(PETSC_HAVE_CUSP) 45 if (Y->valid_GPU_matrix != PETSC_CUSP_UNALLOCATED) { 46 Y->valid_GPU_matrix = PETSC_CUSP_CPU; 47 } 48 #endif 49 #if defined(PETSC_HAVE_VIENNACL) 50 if (Y->valid_GPU_matrix != PETSC_VIENNACL_UNALLOCATED) { 51 Y->valid_GPU_matrix = PETSC_VIENNACL_CPU; 52 } 53 #endif 54 PetscFunctionReturn(0); 55 } 56 57 #undef __FUNCT__ 58 #define __FUNCT__ "MatAXPY_Basic" 59 PetscErrorCode MatAXPY_Basic(Mat Y,PetscScalar a,Mat X,MatStructure str) 60 { 61 PetscInt i,start,end,j,ncols,m,n; 62 PetscErrorCode ierr; 63 const PetscInt *row; 64 PetscScalar *val; 65 const PetscScalar *vals; 66 67 PetscFunctionBegin; 68 ierr = MatGetSize(X,&m,&n);CHKERRQ(ierr); 69 ierr = MatGetOwnershipRange(X,&start,&end);CHKERRQ(ierr); 70 if (a == 1.0) { 71 for (i = start; i < end; i++) { 72 ierr = MatGetRow(X,i,&ncols,&row,&vals);CHKERRQ(ierr); 73 ierr = MatSetValues(Y,1,&i,ncols,row,vals,ADD_VALUES);CHKERRQ(ierr); 74 ierr = MatRestoreRow(X,i,&ncols,&row,&vals);CHKERRQ(ierr); 75 } 76 } else { 77 ierr = PetscMalloc1((n+1),&val);CHKERRQ(ierr); 78 for (i=start; i<end; i++) { 79 ierr = MatGetRow(X,i,&ncols,&row,&vals);CHKERRQ(ierr); 80 for (j=0; j<ncols; j++) { 81 val[j] = a*vals[j]; 82 } 83 ierr = MatSetValues(Y,1,&i,ncols,row,val,ADD_VALUES);CHKERRQ(ierr); 84 ierr = MatRestoreRow(X,i,&ncols,&row,&vals);CHKERRQ(ierr); 85 } 86 ierr = PetscFree(val);CHKERRQ(ierr); 87 } 88 ierr = MatAssemblyBegin(Y,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 89 ierr = MatAssemblyEnd(Y,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 90 PetscFunctionReturn(0); 91 } 92 93 #undef __FUNCT__ 94 #define __FUNCT__ "MatAXPY_BasicWithPreallocation" 95 PetscErrorCode MatAXPY_BasicWithPreallocation(Mat B,Mat Y,PetscScalar a,Mat X,MatStructure str) 96 { 97 PetscInt i,start,end,j,ncols,m,n; 98 PetscErrorCode ierr; 99 const PetscInt *row; 100 PetscScalar *val; 101 const PetscScalar *vals; 102 103 PetscFunctionBegin; 104 ierr = MatGetSize(X,&m,&n);CHKERRQ(ierr); 105 ierr = MatGetOwnershipRange(X,&start,&end);CHKERRQ(ierr); 106 if (a == 1.0) { 107 for (i = start; i < end; i++) { 108 ierr = MatGetRow(Y,i,&ncols,&row,&vals);CHKERRQ(ierr); 109 ierr = MatSetValues(B,1,&i,ncols,row,vals,ADD_VALUES);CHKERRQ(ierr); 110 ierr = MatRestoreRow(Y,i,&ncols,&row,&vals);CHKERRQ(ierr); 111 112 ierr = MatGetRow(X,i,&ncols,&row,&vals);CHKERRQ(ierr); 113 ierr = MatSetValues(B,1,&i,ncols,row,vals,ADD_VALUES);CHKERRQ(ierr); 114 ierr = MatRestoreRow(X,i,&ncols,&row,&vals);CHKERRQ(ierr); 115 } 116 } else { 117 ierr = PetscMalloc1((n+1),&val);CHKERRQ(ierr); 118 for (i=start; i<end; i++) { 119 ierr = MatGetRow(Y,i,&ncols,&row,&vals);CHKERRQ(ierr); 120 ierr = MatSetValues(B,1,&i,ncols,row,vals,ADD_VALUES);CHKERRQ(ierr); 121 ierr = MatRestoreRow(Y,i,&ncols,&row,&vals);CHKERRQ(ierr); 122 123 ierr = MatGetRow(X,i,&ncols,&row,&vals);CHKERRQ(ierr); 124 for (j=0; j<ncols; j++) { 125 val[j] = a*vals[j]; 126 } 127 ierr = MatSetValues(B,1,&i,ncols,row,val,ADD_VALUES);CHKERRQ(ierr); 128 ierr = MatRestoreRow(X,i,&ncols,&row,&vals);CHKERRQ(ierr); 129 } 130 ierr = PetscFree(val);CHKERRQ(ierr); 131 } 132 ierr = MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 133 ierr = MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 134 PetscFunctionReturn(0); 135 } 136 137 #undef __FUNCT__ 138 #define __FUNCT__ "MatShift" 139 /*@ 140 MatShift - Computes Y = Y + a I, where a is a PetscScalar and I is the identity matrix. 141 142 Neighbor-wise Collective on Mat 143 144 Input Parameters: 145 + Y - the matrices 146 - a - the PetscScalar 147 148 Level: intermediate 149 150 .keywords: matrix, add, shift 151 152 .seealso: MatDiagonalSet() 153 @*/ 154 PetscErrorCode MatShift(Mat Y,PetscScalar a) 155 { 156 PetscErrorCode ierr; 157 PetscInt i,start,end; 158 159 PetscFunctionBegin; 160 PetscValidHeaderSpecific(Y,MAT_CLASSID,1); 161 if (!Y->assembled) SETERRQ(PetscObjectComm((PetscObject)Y),PETSC_ERR_ARG_WRONGSTATE,"Not for unassembled matrix"); 162 if (Y->factortype) SETERRQ(PetscObjectComm((PetscObject)Y),PETSC_ERR_ARG_WRONGSTATE,"Not for factored matrix"); 163 MatCheckPreallocated(Y,1); 164 165 if (Y->ops->shift) { 166 ierr = (*Y->ops->shift)(Y,a);CHKERRQ(ierr); 167 } else { 168 PetscScalar alpha = a; 169 ierr = MatGetOwnershipRange(Y,&start,&end);CHKERRQ(ierr); 170 for (i=start; i<end; i++) { 171 ierr = MatSetValues(Y,1,&i,1,&i,&alpha,ADD_VALUES);CHKERRQ(ierr); 172 } 173 ierr = MatAssemblyBegin(Y,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 174 ierr = MatAssemblyEnd(Y,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 175 } 176 #if defined(PETSC_HAVE_CUSP) 177 if (Y->valid_GPU_matrix != PETSC_CUSP_UNALLOCATED) { 178 Y->valid_GPU_matrix = PETSC_CUSP_CPU; 179 } 180 #endif 181 #if defined(PETSC_HAVE_VIENNACL) 182 if (Y->valid_GPU_matrix != PETSC_VIENNACL_UNALLOCATED) { 183 Y->valid_GPU_matrix = PETSC_VIENNACL_CPU; 184 } 185 #endif 186 PetscFunctionReturn(0); 187 } 188 189 #undef __FUNCT__ 190 #define __FUNCT__ "MatDiagonalSet_Default" 191 PetscErrorCode MatDiagonalSet_Default(Mat Y,Vec D,InsertMode is) 192 { 193 PetscErrorCode ierr; 194 PetscInt i,start,end; 195 PetscScalar *v; 196 197 PetscFunctionBegin; 198 ierr = MatGetOwnershipRange(Y,&start,&end);CHKERRQ(ierr); 199 ierr = VecGetArray(D,&v);CHKERRQ(ierr); 200 for (i=start; i<end; i++) { 201 ierr = MatSetValues(Y,1,&i,1,&i,v+i-start,is);CHKERRQ(ierr); 202 } 203 ierr = VecRestoreArray(D,&v);CHKERRQ(ierr); 204 ierr = MatAssemblyBegin(Y,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 205 ierr = MatAssemblyEnd(Y,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 206 PetscFunctionReturn(0); 207 } 208 209 #undef __FUNCT__ 210 #define __FUNCT__ "MatDiagonalSet" 211 /*@ 212 MatDiagonalSet - Computes Y = Y + D, where D is a diagonal matrix 213 that is represented as a vector. Or Y[i,i] = D[i] if InsertMode is 214 INSERT_VALUES. 215 216 Input Parameters: 217 + Y - the input matrix 218 . D - the diagonal matrix, represented as a vector 219 - i - INSERT_VALUES or ADD_VALUES 220 221 Neighbor-wise Collective on Mat and Vec 222 223 Level: intermediate 224 225 .keywords: matrix, add, shift, diagonal 226 227 .seealso: MatShift() 228 @*/ 229 PetscErrorCode MatDiagonalSet(Mat Y,Vec D,InsertMode is) 230 { 231 PetscErrorCode ierr; 232 PetscInt matlocal,veclocal; 233 234 PetscFunctionBegin; 235 PetscValidHeaderSpecific(Y,MAT_CLASSID,1); 236 PetscValidHeaderSpecific(D,VEC_CLASSID,2); 237 ierr = MatGetLocalSize(Y,&matlocal,NULL);CHKERRQ(ierr); 238 ierr = VecGetLocalSize(D,&veclocal);CHKERRQ(ierr); 239 if (matlocal != veclocal) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_INCOMP,"Number local rows of matrix %D does not match that of vector for diagonal %D",matlocal,veclocal); 240 if (Y->ops->diagonalset) { 241 ierr = (*Y->ops->diagonalset)(Y,D,is);CHKERRQ(ierr); 242 } else { 243 ierr = MatDiagonalSet_Default(Y,D,is);CHKERRQ(ierr); 244 } 245 PetscFunctionReturn(0); 246 } 247 248 #undef __FUNCT__ 249 #define __FUNCT__ "MatAYPX" 250 /*@ 251 MatAYPX - Computes Y = a*Y + X. 252 253 Logically on Mat 254 255 Input Parameters: 256 + a - the PetscScalar multiplier 257 . Y - the first matrix 258 . X - the second matrix 259 - str - either SAME_NONZERO_PATTERN, DIFFERENT_NONZERO_PATTERN or SUBSET_NONZERO_PATTERN 260 261 Level: intermediate 262 263 .keywords: matrix, add 264 265 .seealso: MatAXPY() 266 @*/ 267 PetscErrorCode MatAYPX(Mat Y,PetscScalar a,Mat X,MatStructure str) 268 { 269 PetscScalar one = 1.0; 270 PetscErrorCode ierr; 271 PetscInt mX,mY,nX,nY; 272 273 PetscFunctionBegin; 274 PetscValidHeaderSpecific(X,MAT_CLASSID,3); 275 PetscValidHeaderSpecific(Y,MAT_CLASSID,1); 276 PetscValidLogicalCollectiveScalar(Y,a,2); 277 ierr = MatGetSize(X,&mX,&nX);CHKERRQ(ierr); 278 ierr = MatGetSize(X,&mY,&nY);CHKERRQ(ierr); 279 if (mX != mY || nX != nY) SETERRQ4(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Non conforming matrices: %D %D first %D %D second",mX,mY,nX,nY); 280 281 ierr = MatScale(Y,a);CHKERRQ(ierr); 282 ierr = MatAXPY(Y,one,X,str);CHKERRQ(ierr); 283 PetscFunctionReturn(0); 284 } 285 286 #undef __FUNCT__ 287 #define __FUNCT__ "MatComputeExplicitOperator" 288 /*@ 289 MatComputeExplicitOperator - Computes the explicit matrix 290 291 Collective on Mat 292 293 Input Parameter: 294 . inmat - the matrix 295 296 Output Parameter: 297 . mat - the explict preconditioned operator 298 299 Notes: 300 This computation is done by applying the operators to columns of the 301 identity matrix. 302 303 Currently, this routine uses a dense matrix format when 1 processor 304 is used and a sparse format otherwise. This routine is costly in general, 305 and is recommended for use only with relatively small systems. 306 307 Level: advanced 308 309 .keywords: Mat, compute, explicit, operator 310 @*/ 311 PetscErrorCode MatComputeExplicitOperator(Mat inmat,Mat *mat) 312 { 313 Vec in,out; 314 PetscErrorCode ierr; 315 PetscInt i,m,n,M,N,*rows,start,end; 316 MPI_Comm comm; 317 PetscScalar *array,zero = 0.0,one = 1.0; 318 PetscMPIInt size; 319 320 PetscFunctionBegin; 321 PetscValidHeaderSpecific(inmat,MAT_CLASSID,1); 322 PetscValidPointer(mat,2); 323 324 ierr = PetscObjectGetComm((PetscObject)inmat,&comm);CHKERRQ(ierr); 325 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 326 327 ierr = MatGetLocalSize(inmat,&m,&n);CHKERRQ(ierr); 328 ierr = MatGetSize(inmat,&M,&N);CHKERRQ(ierr); 329 ierr = MatCreateVecs(inmat,&in,&out);CHKERRQ(ierr); 330 ierr = VecSetOption(in,VEC_IGNORE_OFF_PROC_ENTRIES,PETSC_TRUE);CHKERRQ(ierr); 331 ierr = VecGetOwnershipRange(out,&start,&end);CHKERRQ(ierr); 332 ierr = PetscMalloc1(m,&rows);CHKERRQ(ierr); 333 for (i=0; i<m; i++) rows[i] = start + i; 334 335 ierr = MatCreate(comm,mat);CHKERRQ(ierr); 336 ierr = MatSetSizes(*mat,m,n,M,N);CHKERRQ(ierr); 337 if (size == 1) { 338 ierr = MatSetType(*mat,MATSEQDENSE);CHKERRQ(ierr); 339 ierr = MatSeqDenseSetPreallocation(*mat,NULL);CHKERRQ(ierr); 340 } else { 341 ierr = MatSetType(*mat,MATMPIAIJ);CHKERRQ(ierr); 342 ierr = MatMPIAIJSetPreallocation(*mat,n,NULL,N-n,NULL);CHKERRQ(ierr); 343 } 344 345 for (i=0; i<N; i++) { 346 347 ierr = VecSet(in,zero);CHKERRQ(ierr); 348 ierr = VecSetValues(in,1,&i,&one,INSERT_VALUES);CHKERRQ(ierr); 349 ierr = VecAssemblyBegin(in);CHKERRQ(ierr); 350 ierr = VecAssemblyEnd(in);CHKERRQ(ierr); 351 352 ierr = MatMult(inmat,in,out);CHKERRQ(ierr); 353 354 ierr = VecGetArray(out,&array);CHKERRQ(ierr); 355 ierr = MatSetValues(*mat,m,rows,1,&i,array,INSERT_VALUES);CHKERRQ(ierr); 356 ierr = VecRestoreArray(out,&array);CHKERRQ(ierr); 357 358 } 359 ierr = PetscFree(rows);CHKERRQ(ierr); 360 ierr = VecDestroy(&out);CHKERRQ(ierr); 361 ierr = VecDestroy(&in);CHKERRQ(ierr); 362 ierr = MatAssemblyBegin(*mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 363 ierr = MatAssemblyEnd(*mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 364 PetscFunctionReturn(0); 365 } 366 367 #undef __FUNCT__ 368 #define __FUNCT__ "MatChop" 369 /*@ 370 MatChop - Set all values in the matrix less than the tolerance to zero 371 372 Input Parameters: 373 + A - The matrix 374 - tol - The zero tolerance 375 376 Output Parameters: 377 . A - The chopped matrix 378 379 Level: intermediate 380 381 .seealso: MatCreate(), MatZeroEntries() 382 @*/ 383 PetscErrorCode MatChop(Mat A, PetscReal tol) 384 { 385 PetscScalar *newVals; 386 PetscInt *newCols; 387 PetscInt rStart, rEnd, numRows, maxRows, r, colMax = 0; 388 PetscErrorCode ierr; 389 390 PetscFunctionBegin; 391 ierr = MatGetOwnershipRange(A, &rStart, &rEnd);CHKERRQ(ierr); 392 for (r = rStart; r < rEnd; ++r) { 393 PetscInt ncols; 394 395 ierr = MatGetRow(A, r, &ncols, NULL, NULL);CHKERRQ(ierr); 396 colMax = PetscMax(colMax, ncols);CHKERRQ(ierr); 397 ierr = MatRestoreRow(A, r, &ncols, NULL, NULL);CHKERRQ(ierr); 398 } 399 numRows = rEnd - rStart; 400 ierr = MPI_Allreduce(&numRows, &maxRows, 1, MPIU_INT, MPI_MAX, PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 401 ierr = PetscMalloc2(colMax,&newCols,colMax,&newVals);CHKERRQ(ierr); 402 for (r = rStart; r < rStart+maxRows; ++r) { 403 const PetscScalar *vals; 404 const PetscInt *cols; 405 PetscInt ncols, newcols, c; 406 407 if (r < rEnd) { 408 ierr = MatGetRow(A, r, &ncols, &cols, &vals);CHKERRQ(ierr); 409 for (c = 0; c < ncols; ++c) { 410 newCols[c] = cols[c]; 411 newVals[c] = PetscAbsScalar(vals[c]) < tol ? 0.0 : vals[c]; 412 } 413 newcols = ncols; 414 ierr = MatRestoreRow(A, r, &ncols, &cols, &vals);CHKERRQ(ierr); 415 ierr = MatSetValues(A, 1, &r, newcols, newCols, newVals, INSERT_VALUES);CHKERRQ(ierr); 416 } 417 ierr = MatAssemblyBegin(A, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 418 ierr = MatAssemblyEnd(A, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 419 } 420 ierr = PetscFree2(newCols,newVals);CHKERRQ(ierr); 421 PetscFunctionReturn(0); 422 } 423