1 2 #if !defined(__AIJ_H) 3 #define __AIJ_H 4 5 #include <petsc/private/matimpl.h> 6 #include <petscctable.h> 7 8 /* Operations provided by MATSEQAIJ and its subclasses */ 9 typedef struct { 10 PetscErrorCode (*getarray)(Mat,PetscScalar **); 11 PetscErrorCode (*restorearray)(Mat,PetscScalar **); 12 PetscErrorCode (*getarrayread)(Mat,const PetscScalar **); 13 PetscErrorCode (*restorearrayread)(Mat,const PetscScalar **); 14 PetscErrorCode (*getarraywrite)(Mat,PetscScalar **); 15 PetscErrorCode (*restorearraywrite)(Mat,PetscScalar **); 16 } Mat_SeqAIJOps; 17 18 /* 19 Struct header shared by SeqAIJ, SeqBAIJ and SeqSBAIJ matrix formats 20 */ 21 #define SEQAIJHEADER(datatype) \ 22 PetscBool roworiented; /* if true, row-oriented input, default */ \ 23 PetscInt nonew; /* 1 don't add new nonzeros, -1 generate error on new */ \ 24 PetscInt nounused; /* -1 generate error on unused space */ \ 25 PetscBool singlemalloc; /* if true a, i, and j have been obtained with one big malloc */ \ 26 PetscInt maxnz; /* allocated nonzeros */ \ 27 PetscInt *imax; /* maximum space allocated for each row */ \ 28 PetscInt *ilen; /* actual length of each row */ \ 29 PetscInt *ipre; /* space preallocated for each row by user */ \ 30 PetscBool free_imax_ilen; \ 31 PetscInt reallocs; /* number of mallocs done during MatSetValues() \ 32 as more values are set than were prealloced */\ 33 PetscInt rmax; /* max nonzeros in any row */ \ 34 PetscBool keepnonzeropattern; /* keeps matrix structure same in calls to MatZeroRows()*/ \ 35 PetscBool ignorezeroentries; \ 36 PetscBool free_ij; /* free the column indices j and row offsets i when the matrix is destroyed */ \ 37 PetscBool free_a; /* free the numerical values when matrix is destroy */ \ 38 Mat_CompressedRow compressedrow; /* use compressed row format */ \ 39 PetscInt nz; /* nonzeros */ \ 40 PetscInt *i; /* pointer to beginning of each row */ \ 41 PetscInt *j; /* column values: j + i[k] - 1 is start of row k */ \ 42 PetscInt *diag; /* pointers to diagonal elements */ \ 43 PetscInt nonzerorowcnt; /* how many rows have nonzero entries */ \ 44 PetscBool free_diag; \ 45 datatype *a; /* nonzero elements */ \ 46 PetscScalar *solve_work; /* work space used in MatSolve */ \ 47 IS row, col, icol; /* index sets, used for reorderings */ \ 48 PetscBool pivotinblocks; /* pivot inside factorization of each diagonal block */ \ 49 Mat parent; /* set if this matrix was formed with MatDuplicate(...,MAT_SHARE_NONZERO_PATTERN,....); \ 50 means that this shares some data structures with the parent including diag, ilen, imax, i, j */\ 51 Mat_SubSppt *submatis1; /* used by MatCreateSubMatrices_MPIXAIJ_Local */ \ 52 Mat_SeqAIJOps ops[1] /* operations for SeqAIJ and its subclasses */ 53 54 typedef struct { 55 MatTransposeColoring matcoloring; 56 Mat Bt_den; /* dense matrix of B^T */ 57 Mat ABt_den; /* dense matrix of A*B^T */ 58 PetscBool usecoloring; 59 } Mat_MatMatTransMult; 60 61 typedef struct { /* used by MatTransposeMatMult() */ 62 Mat At; /* transpose of the first matrix */ 63 Mat mA; /* maij matrix of A */ 64 Vec bt,ct; /* vectors to hold locally transposed arrays of B and C */ 65 PetscBool updateAt; /* flg to avoid recomputing At in MatProductNumeric_AtB_SeqAIJ_SeqAIJ() */ 66 /* used by PtAP */ 67 void *data; 68 PetscErrorCode (*destroy)(void*); 69 } Mat_MatTransMatMult; 70 71 typedef struct { 72 PetscInt *api,*apj; /* symbolic structure of A*P */ 73 PetscScalar *apa; /* temporary array for storing one row of A*P */ 74 } Mat_AP; 75 76 typedef struct { 77 MatTransposeColoring matcoloring; 78 Mat Rt; /* sparse or dense matrix of R^T */ 79 Mat RARt; /* dense matrix of R*A*R^T */ 80 Mat ARt; /* A*R^T used for the case -matrart_color_art */ 81 MatScalar *work; /* work array to store columns of A*R^T used in MatMatMatMultNumeric_SeqAIJ_SeqAIJ_SeqDense() */ 82 /* free intermediate products needed for PtAP */ 83 void *data; 84 PetscErrorCode (*destroy)(void*); 85 } Mat_RARt; 86 87 typedef struct { 88 Mat BC; /* temp matrix for storing B*C */ 89 } Mat_MatMatMatMult; 90 91 /* 92 MATSEQAIJ format - Compressed row storage (also called Yale sparse matrix 93 format) or compressed sparse row (CSR). The i[] and j[] arrays start at 0. For example, 94 j[i[k]+p] is the pth column in row k. Note that the diagonal 95 matrix elements are stored with the rest of the nonzeros (not separately). 96 */ 97 98 /* Info about i-nodes (identical nodes) helper class for SeqAIJ */ 99 typedef struct { 100 MatScalar *bdiag,*ibdiag,*ssor_work; /* diagonal blocks of matrix used for MatSOR_SeqAIJ_Inode() */ 101 PetscInt bdiagsize; /* length of bdiag and ibdiag */ 102 PetscBool ibdiagvalid; /* do ibdiag[] and bdiag[] contain the most recent values */ 103 104 PetscBool use; 105 PetscInt node_count; /* number of inodes */ 106 PetscInt *size; /* size of each inode */ 107 PetscInt limit; /* inode limit */ 108 PetscInt max_limit; /* maximum supported inode limit */ 109 PetscBool checked; /* if inodes have been checked for */ 110 PetscObjectState mat_nonzerostate; /* non-zero state when inodes were checked for */ 111 } Mat_SeqAIJ_Inode; 112 113 PETSC_INTERN PetscErrorCode MatView_SeqAIJ_Inode(Mat,PetscViewer); 114 PETSC_INTERN PetscErrorCode MatAssemblyEnd_SeqAIJ_Inode(Mat,MatAssemblyType); 115 PETSC_INTERN PetscErrorCode MatDestroy_SeqAIJ_Inode(Mat); 116 PETSC_INTERN PetscErrorCode MatCreate_SeqAIJ_Inode(Mat); 117 PETSC_INTERN PetscErrorCode MatSetOption_SeqAIJ_Inode(Mat,MatOption,PetscBool); 118 PETSC_INTERN PetscErrorCode MatDuplicate_SeqAIJ_Inode(Mat,MatDuplicateOption,Mat*); 119 PETSC_INTERN PetscErrorCode MatDuplicateNoCreate_SeqAIJ(Mat,Mat,MatDuplicateOption,PetscBool); 120 PETSC_INTERN PetscErrorCode MatLUFactorNumeric_SeqAIJ_Inode_inplace(Mat,Mat,const MatFactorInfo*); 121 PETSC_INTERN PetscErrorCode MatLUFactorNumeric_SeqAIJ_Inode(Mat,Mat,const MatFactorInfo*); 122 PETSC_INTERN PetscErrorCode MatSeqAIJGetArray_SeqAIJ(Mat,PetscScalar**); 123 PETSC_INTERN PetscErrorCode MatSeqAIJRestoreArray_SeqAIJ(Mat,PetscScalar**); 124 125 typedef struct { 126 SEQAIJHEADER(MatScalar); 127 Mat_SeqAIJ_Inode inode; 128 MatScalar *saved_values; /* location for stashing nonzero values of matrix */ 129 130 PetscScalar *idiag,*mdiag,*ssor_work; /* inverse of diagonal entries, diagonal values and workspace for Eisenstat trick */ 131 PetscBool idiagvalid; /* current idiag[] and mdiag[] are valid */ 132 PetscScalar *ibdiag; /* inverses of block diagonals */ 133 PetscBool ibdiagvalid; /* inverses of block diagonals are valid. */ 134 PetscBool diagonaldense; /* all entries along the diagonal have been set; i.e. no missing diagonal terms */ 135 PetscScalar fshift,omega; /* last used omega and fshift */ 136 137 /* MatSetValuesCOO() related fields on host */ 138 PetscCount coo_n; /* Number of entries in MatSetPreallocationCOO() */ 139 PetscCount Atot; /* Total number of valid (i.e., w/ non-negative indices) entries in the COO array */ 140 PetscCount *jmap; /* perm[jmap[i]..jmap[i+1]) give indices of entries in v[] associated with i-th nonzero of the matrix */ 141 PetscCount *perm; /* The permutation array in sorting (i,j) by row and then by col */ 142 } Mat_SeqAIJ; 143 144 /* 145 Frees the a, i, and j arrays from the XAIJ (AIJ, BAIJ, and SBAIJ) matrix types 146 */ 147 static inline PetscErrorCode MatSeqXAIJFreeAIJ(Mat AA,MatScalar **a,PetscInt **j,PetscInt **i) 148 { 149 Mat_SeqAIJ *A = (Mat_SeqAIJ*) AA->data; 150 if (A->singlemalloc) { 151 PetscCall(PetscFree3(*a,*j,*i)); 152 } else { 153 if (A->free_a) PetscCall(PetscFree(*a)); 154 if (A->free_ij) PetscCall(PetscFree(*j)); 155 if (A->free_ij) PetscCall(PetscFree(*i)); 156 } 157 return 0; 158 } 159 /* 160 Allocates larger a, i, and j arrays for the XAIJ (AIJ, BAIJ, and SBAIJ) matrix types 161 This is a macro because it takes the datatype as an argument which can be either a Mat or a MatScalar 162 */ 163 #define MatSeqXAIJReallocateAIJ(Amat,AM,BS2,NROW,ROW,COL,RMAX,AA,AI,AJ,RP,AP,AIMAX,NONEW,datatype) \ 164 if (NROW >= RMAX) { \ 165 Mat_SeqAIJ *Ain = (Mat_SeqAIJ*)Amat->data; \ 166 /* there is no extra room in row, therefore enlarge */ \ 167 PetscInt CHUNKSIZE = 15,new_nz = AI[AM] + CHUNKSIZE,len,*new_i=NULL,*new_j=NULL; \ 168 datatype *new_a; \ 169 \ 170 PetscCheckFalse(NONEW == -2,PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"New nonzero at (%" PetscInt_FMT ",%" PetscInt_FMT ") caused a malloc\nUse MatSetOption(A, MAT_NEW_NONZERO_ALLOCATION_ERR, PETSC_FALSE) to turn off this check",ROW,COL); \ 171 /* malloc new storage space */ \ 172 PetscCall(PetscMalloc3(BS2*new_nz,&new_a,new_nz,&new_j,AM+1,&new_i)); \ 173 \ 174 /* copy over old data into new slots */ \ 175 for (ii=0; ii<ROW+1; ii++) {new_i[ii] = AI[ii];} \ 176 for (ii=ROW+1; ii<AM+1; ii++) {new_i[ii] = AI[ii]+CHUNKSIZE;} \ 177 PetscCall(PetscArraycpy(new_j,AJ,AI[ROW]+NROW)); \ 178 len = (new_nz - CHUNKSIZE - AI[ROW] - NROW); \ 179 PetscCall(PetscArraycpy(new_j+AI[ROW]+NROW+CHUNKSIZE,AJ+AI[ROW]+NROW,len)); \ 180 PetscCall(PetscArraycpy(new_a,AA,BS2*(AI[ROW]+NROW))); \ 181 PetscCall(PetscArrayzero(new_a+BS2*(AI[ROW]+NROW),BS2*CHUNKSIZE)); \ 182 PetscCall(PetscArraycpy(new_a+BS2*(AI[ROW]+NROW+CHUNKSIZE),AA+BS2*(AI[ROW]+NROW),BS2*len)); \ 183 /* free up old matrix storage */ \ 184 PetscCall(MatSeqXAIJFreeAIJ(A,&Ain->a,&Ain->j,&Ain->i)); \ 185 AA = new_a; \ 186 Ain->a = (MatScalar*) new_a; \ 187 AI = Ain->i = new_i; AJ = Ain->j = new_j; \ 188 Ain->singlemalloc = PETSC_TRUE; \ 189 \ 190 RP = AJ + AI[ROW]; AP = AA + BS2*AI[ROW]; \ 191 RMAX = AIMAX[ROW] = AIMAX[ROW] + CHUNKSIZE; \ 192 Ain->maxnz += BS2*CHUNKSIZE; \ 193 Ain->reallocs++; \ 194 } \ 195 196 #define MatSeqXAIJReallocateAIJ_structure_only(Amat,AM,BS2,NROW,ROW,COL,RMAX,AI,AJ,RP,AIMAX,NONEW,datatype) \ 197 if (NROW >= RMAX) { \ 198 Mat_SeqAIJ *Ain = (Mat_SeqAIJ*)Amat->data; \ 199 /* there is no extra room in row, therefore enlarge */ \ 200 PetscInt CHUNKSIZE = 15,new_nz = AI[AM] + CHUNKSIZE,len,*new_i=NULL,*new_j=NULL; \ 201 \ 202 PetscCheckFalse(NONEW == -2,PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"New nonzero at (%" PetscInt_FMT ",%" PetscInt_FMT ") caused a malloc\nUse MatSetOption(A, MAT_NEW_NONZERO_ALLOCATION_ERR, PETSC_FALSE) to turn off this check",ROW,COL); \ 203 /* malloc new storage space */ \ 204 PetscCall(PetscMalloc1(new_nz,&new_j)); \ 205 PetscCall(PetscMalloc1(AM+1,&new_i));\ 206 \ 207 /* copy over old data into new slots */ \ 208 for (ii=0; ii<ROW+1; ii++) {new_i[ii] = AI[ii];} \ 209 for (ii=ROW+1; ii<AM+1; ii++) {new_i[ii] = AI[ii]+CHUNKSIZE;} \ 210 PetscCall(PetscArraycpy(new_j,AJ,AI[ROW]+NROW)); \ 211 len = (new_nz - CHUNKSIZE - AI[ROW] - NROW); \ 212 PetscCall(PetscArraycpy(new_j+AI[ROW]+NROW+CHUNKSIZE,AJ+AI[ROW]+NROW,len)); \ 213 \ 214 /* free up old matrix storage */ \ 215 PetscCall(MatSeqXAIJFreeAIJ(A,&Ain->a,&Ain->j,&Ain->i)); \ 216 Ain->a = NULL; \ 217 AI = Ain->i = new_i; AJ = Ain->j = new_j; \ 218 Ain->singlemalloc = PETSC_FALSE; \ 219 Ain->free_a = PETSC_FALSE; \ 220 \ 221 RP = AJ + AI[ROW]; \ 222 RMAX = AIMAX[ROW] = AIMAX[ROW] + CHUNKSIZE; \ 223 Ain->maxnz += BS2*CHUNKSIZE; \ 224 Ain->reallocs++; \ 225 } \ 226 227 PETSC_INTERN PetscErrorCode MatSeqAIJSetPreallocation_SeqAIJ(Mat,PetscInt,const PetscInt*); 228 PETSC_INTERN PetscErrorCode MatSetPreallocationCOO_SeqAIJ(Mat,PetscCount,const PetscInt[],const PetscInt[]); 229 PETSC_INTERN PetscErrorCode MatResetPreallocationCOO_SeqAIJ(Mat); 230 231 PETSC_INTERN PetscErrorCode MatILUFactorSymbolic_SeqAIJ_inplace(Mat,Mat,IS,IS,const MatFactorInfo*); 232 PETSC_INTERN PetscErrorCode MatILUFactorSymbolic_SeqAIJ(Mat,Mat,IS,IS,const MatFactorInfo*); 233 PETSC_INTERN PetscErrorCode MatILUFactorSymbolic_SeqAIJ_ilu0(Mat,Mat,IS,IS,const MatFactorInfo*); 234 235 PETSC_INTERN PetscErrorCode MatICCFactorSymbolic_SeqAIJ_inplace(Mat,Mat,IS,const MatFactorInfo*); 236 PETSC_INTERN PetscErrorCode MatICCFactorSymbolic_SeqAIJ(Mat,Mat,IS,const MatFactorInfo*); 237 PETSC_INTERN PetscErrorCode MatCholeskyFactorSymbolic_SeqAIJ_inplace(Mat,Mat,IS,const MatFactorInfo*); 238 PETSC_INTERN PetscErrorCode MatCholeskyFactorSymbolic_SeqAIJ(Mat,Mat,IS,const MatFactorInfo*); 239 PETSC_INTERN PetscErrorCode MatCholeskyFactorNumeric_SeqAIJ_inplace(Mat,Mat,const MatFactorInfo*); 240 PETSC_INTERN PetscErrorCode MatCholeskyFactorNumeric_SeqAIJ(Mat,Mat,const MatFactorInfo*); 241 PETSC_INTERN PetscErrorCode MatDuplicate_SeqAIJ(Mat,MatDuplicateOption,Mat*); 242 PETSC_INTERN PetscErrorCode MatCopy_SeqAIJ(Mat,Mat,MatStructure); 243 PETSC_INTERN PetscErrorCode MatMissingDiagonal_SeqAIJ(Mat,PetscBool*,PetscInt*); 244 PETSC_INTERN PetscErrorCode MatMarkDiagonal_SeqAIJ(Mat); 245 PETSC_INTERN PetscErrorCode MatFindZeroDiagonals_SeqAIJ_Private(Mat,PetscInt*,PetscInt**); 246 247 PETSC_INTERN PetscErrorCode MatMult_SeqAIJ(Mat,Vec,Vec); 248 PETSC_INTERN PetscErrorCode MatMult_SeqAIJ_Inode(Mat,Vec,Vec); 249 PETSC_INTERN PetscErrorCode MatMultAdd_SeqAIJ(Mat,Vec,Vec,Vec); 250 PETSC_INTERN PetscErrorCode MatMultAdd_SeqAIJ_Inode(Mat,Vec,Vec,Vec); 251 PETSC_INTERN PetscErrorCode MatMultTranspose_SeqAIJ(Mat,Vec,Vec); 252 PETSC_INTERN PetscErrorCode MatMultTransposeAdd_SeqAIJ(Mat,Vec,Vec,Vec); 253 PETSC_INTERN PetscErrorCode MatSOR_SeqAIJ(Mat,Vec,PetscReal,MatSORType,PetscReal,PetscInt,PetscInt,Vec); 254 PETSC_INTERN PetscErrorCode MatSOR_SeqAIJ_Inode(Mat,Vec,PetscReal,MatSORType,PetscReal,PetscInt,PetscInt,Vec); 255 256 PETSC_INTERN PetscErrorCode MatSetOption_SeqAIJ(Mat,MatOption,PetscBool); 257 258 PETSC_INTERN PetscErrorCode MatGetSymbolicTranspose_SeqAIJ(Mat,PetscInt *[],PetscInt *[]); 259 PETSC_INTERN PetscErrorCode MatGetSymbolicTransposeReduced_SeqAIJ(Mat,PetscInt,PetscInt,PetscInt *[],PetscInt *[]); 260 PETSC_INTERN PetscErrorCode MatRestoreSymbolicTranspose_SeqAIJ(Mat,PetscInt *[],PetscInt *[]); 261 PETSC_INTERN PetscErrorCode MatTransposeSymbolic_SeqAIJ(Mat,Mat*); 262 PETSC_INTERN PetscErrorCode MatTranspose_SeqAIJ(Mat,MatReuse,Mat*); 263 PETSC_INTERN PetscErrorCode MatToSymmetricIJ_SeqAIJ(PetscInt,PetscInt*,PetscInt*,PetscBool,PetscInt,PetscInt,PetscInt**,PetscInt**); 264 PETSC_INTERN PetscErrorCode MatLUFactorSymbolic_SeqAIJ_inplace(Mat,Mat,IS,IS,const MatFactorInfo*); 265 PETSC_INTERN PetscErrorCode MatLUFactorSymbolic_SeqAIJ(Mat,Mat,IS,IS,const MatFactorInfo*); 266 PETSC_INTERN PetscErrorCode MatLUFactorNumeric_SeqAIJ_inplace(Mat,Mat,const MatFactorInfo*); 267 PETSC_INTERN PetscErrorCode MatLUFactorNumeric_SeqAIJ(Mat,Mat,const MatFactorInfo*); 268 PETSC_INTERN PetscErrorCode MatLUFactorNumeric_SeqAIJ_InplaceWithPerm(Mat,Mat,const MatFactorInfo*); 269 PETSC_INTERN PetscErrorCode MatLUFactor_SeqAIJ(Mat,IS,IS,const MatFactorInfo*); 270 PETSC_INTERN PetscErrorCode MatSolve_SeqAIJ_inplace(Mat,Vec,Vec); 271 PETSC_INTERN PetscErrorCode MatSolve_SeqAIJ(Mat,Vec,Vec); 272 PETSC_INTERN PetscErrorCode MatSolve_SeqAIJ_Inode_inplace(Mat,Vec,Vec); 273 PETSC_INTERN PetscErrorCode MatSolve_SeqAIJ_Inode(Mat,Vec,Vec); 274 PETSC_INTERN PetscErrorCode MatSolve_SeqAIJ_NaturalOrdering_inplace(Mat,Vec,Vec); 275 PETSC_INTERN PetscErrorCode MatSolve_SeqAIJ_NaturalOrdering(Mat,Vec,Vec); 276 PETSC_INTERN PetscErrorCode MatSolve_SeqAIJ_InplaceWithPerm(Mat,Vec,Vec); 277 PETSC_INTERN PetscErrorCode MatSolveAdd_SeqAIJ_inplace(Mat,Vec,Vec,Vec); 278 PETSC_INTERN PetscErrorCode MatSolveAdd_SeqAIJ(Mat,Vec,Vec,Vec); 279 PETSC_INTERN PetscErrorCode MatSolveTranspose_SeqAIJ_inplace(Mat,Vec,Vec); 280 PETSC_INTERN PetscErrorCode MatSolveTranspose_SeqAIJ(Mat,Vec,Vec); 281 PETSC_INTERN PetscErrorCode MatSolveTransposeAdd_SeqAIJ_inplace(Mat,Vec,Vec,Vec); 282 PETSC_INTERN PetscErrorCode MatSolveTransposeAdd_SeqAIJ(Mat,Vec,Vec,Vec); 283 PETSC_INTERN PetscErrorCode MatMatSolve_SeqAIJ_inplace(Mat,Mat,Mat); 284 PETSC_INTERN PetscErrorCode MatMatSolve_SeqAIJ(Mat,Mat,Mat); 285 PETSC_INTERN PetscErrorCode MatEqual_SeqAIJ(Mat,Mat,PetscBool*); 286 PETSC_INTERN PetscErrorCode MatFDColoringCreate_SeqXAIJ(Mat,ISColoring,MatFDColoring); 287 PETSC_INTERN PetscErrorCode MatFDColoringSetUp_SeqXAIJ(Mat,ISColoring,MatFDColoring); 288 PETSC_INTERN PetscErrorCode MatFDColoringSetUpBlocked_AIJ_Private(Mat,MatFDColoring,PetscInt); 289 PETSC_INTERN PetscErrorCode MatLoad_AIJ_HDF5(Mat,PetscViewer); 290 PETSC_INTERN PetscErrorCode MatLoad_SeqAIJ_Binary(Mat,PetscViewer); 291 PETSC_INTERN PetscErrorCode MatLoad_SeqAIJ(Mat,PetscViewer); 292 PETSC_INTERN PetscErrorCode RegisterApplyPtAPRoutines_Private(Mat); 293 294 #if defined(PETSC_HAVE_HYPRE) 295 PETSC_INTERN PetscErrorCode MatProductSetFromOptions_Transpose_AIJ_AIJ(Mat); 296 #endif 297 PETSC_INTERN PetscErrorCode MatProductSetFromOptions_SeqAIJ(Mat); 298 299 PETSC_INTERN PetscErrorCode MatProductSymbolic_SeqAIJ_SeqAIJ(Mat); 300 PETSC_INTERN PetscErrorCode MatProductSymbolic_PtAP_SeqAIJ_SeqAIJ(Mat); 301 PETSC_INTERN PetscErrorCode MatProductSymbolic_RARt_SeqAIJ_SeqAIJ(Mat); 302 303 PETSC_INTERN PetscErrorCode MatMatMultSymbolic_SeqAIJ_SeqAIJ(Mat,Mat,PetscReal,Mat); 304 PETSC_INTERN PetscErrorCode MatMatMultSymbolic_SeqAIJ_SeqAIJ_Sorted(Mat,Mat,PetscReal,Mat); 305 PETSC_INTERN PetscErrorCode MatMatMultSymbolic_SeqDense_SeqAIJ(Mat,Mat,PetscReal,Mat); 306 PETSC_INTERN PetscErrorCode MatMatMultSymbolic_SeqAIJ_SeqAIJ_Scalable(Mat,Mat,PetscReal,Mat); 307 PETSC_INTERN PetscErrorCode MatMatMultSymbolic_SeqAIJ_SeqAIJ_Scalable_fast(Mat,Mat,PetscReal,Mat); 308 PETSC_INTERN PetscErrorCode MatMatMultSymbolic_SeqAIJ_SeqAIJ_Heap(Mat,Mat,PetscReal,Mat); 309 PETSC_INTERN PetscErrorCode MatMatMultSymbolic_SeqAIJ_SeqAIJ_BTHeap(Mat,Mat,PetscReal,Mat); 310 PETSC_INTERN PetscErrorCode MatMatMultSymbolic_SeqAIJ_SeqAIJ_RowMerge(Mat,Mat,PetscReal,Mat); 311 PETSC_INTERN PetscErrorCode MatMatMultSymbolic_SeqAIJ_SeqAIJ_LLCondensed(Mat,Mat,PetscReal,Mat); 312 #if defined(PETSC_HAVE_HYPRE) 313 PETSC_INTERN PetscErrorCode MatMatMultSymbolic_AIJ_AIJ_wHYPRE(Mat,Mat,PetscReal,Mat); 314 #endif 315 316 PETSC_INTERN PetscErrorCode MatMatMultNumeric_SeqAIJ_SeqAIJ(Mat,Mat,Mat); 317 PETSC_INTERN PetscErrorCode MatMatMultNumeric_SeqAIJ_SeqAIJ_Sorted(Mat,Mat,Mat); 318 319 PETSC_INTERN PetscErrorCode MatMatMultNumeric_SeqDense_SeqAIJ(Mat,Mat,Mat); 320 PETSC_INTERN PetscErrorCode MatMatMultNumeric_SeqAIJ_SeqAIJ_Scalable(Mat,Mat,Mat); 321 322 PETSC_INTERN PetscErrorCode MatPtAPSymbolic_SeqAIJ_SeqAIJ_SparseAxpy(Mat,Mat,PetscReal,Mat); 323 PETSC_INTERN PetscErrorCode MatPtAPNumeric_SeqAIJ_SeqAIJ(Mat,Mat,Mat); 324 PETSC_INTERN PetscErrorCode MatPtAPNumeric_SeqAIJ_SeqAIJ_SparseAxpy(Mat,Mat,Mat); 325 326 PETSC_INTERN PetscErrorCode MatRARtSymbolic_SeqAIJ_SeqAIJ(Mat,Mat,PetscReal,Mat); 327 PETSC_INTERN PetscErrorCode MatRARtSymbolic_SeqAIJ_SeqAIJ_matmattransposemult(Mat,Mat,PetscReal,Mat); 328 PETSC_INTERN PetscErrorCode MatRARtSymbolic_SeqAIJ_SeqAIJ_colorrart(Mat,Mat,PetscReal,Mat); 329 PETSC_INTERN PetscErrorCode MatRARtNumeric_SeqAIJ_SeqAIJ(Mat,Mat,Mat); 330 PETSC_INTERN PetscErrorCode MatRARtNumeric_SeqAIJ_SeqAIJ_matmattransposemult(Mat,Mat,Mat); 331 PETSC_INTERN PetscErrorCode MatRARtNumeric_SeqAIJ_SeqAIJ_colorrart(Mat,Mat,Mat); 332 333 PETSC_INTERN PetscErrorCode MatTransposeMatMultSymbolic_SeqAIJ_SeqAIJ(Mat,Mat,PetscReal,Mat); 334 PETSC_INTERN PetscErrorCode MatTransposeMatMultNumeric_SeqAIJ_SeqAIJ(Mat,Mat,Mat); 335 PETSC_INTERN PetscErrorCode MatDestroy_SeqAIJ_MatTransMatMult(void*); 336 337 PETSC_INTERN PetscErrorCode MatMatTransposeMultSymbolic_SeqAIJ_SeqAIJ(Mat,Mat,PetscReal,Mat); 338 PETSC_INTERN PetscErrorCode MatMatTransposeMultNumeric_SeqAIJ_SeqAIJ(Mat,Mat,Mat); 339 PETSC_INTERN PetscErrorCode MatTransposeColoringCreate_SeqAIJ(Mat,ISColoring,MatTransposeColoring); 340 PETSC_INTERN PetscErrorCode MatTransColoringApplySpToDen_SeqAIJ(MatTransposeColoring,Mat,Mat); 341 PETSC_INTERN PetscErrorCode MatTransColoringApplyDenToSp_SeqAIJ(MatTransposeColoring,Mat,Mat); 342 343 PETSC_INTERN PetscErrorCode MatMatMatMultSymbolic_SeqAIJ_SeqAIJ_SeqAIJ(Mat,Mat,Mat,PetscReal,Mat); 344 PETSC_INTERN PetscErrorCode MatMatMatMultNumeric_SeqAIJ_SeqAIJ_SeqAIJ(Mat,Mat,Mat,Mat); 345 346 PETSC_INTERN PetscErrorCode MatSetRandomSkipColumnRange_SeqAIJ_Private(Mat,PetscInt,PetscInt,PetscRandom); 347 PETSC_INTERN PetscErrorCode MatSetValues_SeqAIJ(Mat,PetscInt,const PetscInt[],PetscInt,const PetscInt[],const PetscScalar[],InsertMode); 348 PETSC_INTERN PetscErrorCode MatGetRow_SeqAIJ(Mat,PetscInt,PetscInt*,PetscInt**,PetscScalar**); 349 PETSC_INTERN PetscErrorCode MatRestoreRow_SeqAIJ(Mat,PetscInt,PetscInt*,PetscInt**,PetscScalar**); 350 PETSC_INTERN PetscErrorCode MatScale_SeqAIJ(Mat,PetscScalar); 351 PETSC_INTERN PetscErrorCode MatDiagonalScale_SeqAIJ(Mat,Vec,Vec); 352 PETSC_INTERN PetscErrorCode MatDiagonalSet_SeqAIJ(Mat,Vec,InsertMode); 353 PETSC_INTERN PetscErrorCode MatAXPY_SeqAIJ(Mat,PetscScalar,Mat,MatStructure); 354 PETSC_INTERN PetscErrorCode MatGetRowIJ_SeqAIJ(Mat,PetscInt,PetscBool,PetscBool,PetscInt*,const PetscInt *[],const PetscInt *[],PetscBool*); 355 PETSC_INTERN PetscErrorCode MatRestoreRowIJ_SeqAIJ(Mat,PetscInt,PetscBool,PetscBool,PetscInt*,const PetscInt *[],const PetscInt *[],PetscBool*); 356 PETSC_INTERN PetscErrorCode MatGetColumnIJ_SeqAIJ(Mat,PetscInt,PetscBool,PetscBool,PetscInt*,const PetscInt *[],const PetscInt *[],PetscBool*); 357 PETSC_INTERN PetscErrorCode MatRestoreColumnIJ_SeqAIJ(Mat,PetscInt,PetscBool,PetscBool,PetscInt*,const PetscInt *[],const PetscInt *[],PetscBool*); 358 PETSC_INTERN PetscErrorCode MatGetColumnIJ_SeqAIJ_Color(Mat,PetscInt,PetscBool,PetscBool,PetscInt*,const PetscInt *[],const PetscInt *[],PetscInt *[],PetscBool*); 359 PETSC_INTERN PetscErrorCode MatRestoreColumnIJ_SeqAIJ_Color(Mat,PetscInt,PetscBool,PetscBool,PetscInt*,const PetscInt *[],const PetscInt *[],PetscInt *[],PetscBool*); 360 PETSC_INTERN PetscErrorCode MatDestroy_SeqAIJ(Mat); 361 PETSC_INTERN PetscErrorCode MatSetUp_SeqAIJ(Mat); 362 PETSC_INTERN PetscErrorCode MatView_SeqAIJ(Mat,PetscViewer); 363 364 PETSC_INTERN PetscErrorCode MatSeqAIJInvalidateDiagonal(Mat); 365 PETSC_INTERN PetscErrorCode MatSeqAIJInvalidateDiagonal_Inode(Mat); 366 PETSC_INTERN PetscErrorCode MatSeqAIJCheckInode(Mat); 367 PETSC_INTERN PetscErrorCode MatSeqAIJCheckInode_FactorLU(Mat); 368 369 PETSC_INTERN PetscErrorCode MatAXPYGetPreallocation_SeqAIJ(Mat,Mat,PetscInt*); 370 371 #if defined(PETSC_HAVE_MATLAB_ENGINE) 372 PETSC_EXTERN PetscErrorCode MatlabEnginePut_SeqAIJ(PetscObject,void*); 373 PETSC_EXTERN PetscErrorCode MatlabEngineGet_SeqAIJ(PetscObject,void*); 374 #endif 375 PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqSBAIJ(Mat,MatType,MatReuse,Mat*); 376 PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqBAIJ(Mat,MatType,MatReuse,Mat*); 377 PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqDense(Mat,MatType,MatReuse,Mat*); 378 PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqAIJCRL(Mat,MatType,MatReuse,Mat*); 379 PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_Elemental(Mat,MatType,MatReuse,Mat*); 380 #if defined(PETSC_HAVE_SCALAPACK) 381 PETSC_INTERN PetscErrorCode MatConvert_AIJ_ScaLAPACK(Mat,MatType,MatReuse,Mat*); 382 #endif 383 PETSC_INTERN PetscErrorCode MatConvert_AIJ_HYPRE(Mat,MatType,MatReuse,Mat*); 384 PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqAIJPERM(Mat,MatType,MatReuse,Mat*); 385 PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqAIJSELL(Mat,MatType,MatReuse,Mat*); 386 PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqAIJMKL(Mat,MatType,MatReuse,Mat*); 387 PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqAIJViennaCL(Mat,MatType,MatReuse,Mat*); 388 PETSC_INTERN PetscErrorCode MatReorderForNonzeroDiagonal_SeqAIJ(Mat,PetscReal,IS,IS); 389 PETSC_INTERN PetscErrorCode MatRARt_SeqAIJ_SeqAIJ(Mat,Mat,MatReuse,PetscReal,Mat*); 390 PETSC_EXTERN PetscErrorCode MatCreate_SeqAIJ(Mat); 391 PETSC_INTERN PetscErrorCode MatAssemblyEnd_SeqAIJ(Mat,MatAssemblyType); 392 PETSC_EXTERN PetscErrorCode MatZeroEntries_SeqAIJ(Mat); 393 394 PETSC_INTERN PetscErrorCode MatAXPYGetPreallocation_SeqX_private(PetscInt,const PetscInt*,const PetscInt*,const PetscInt*,const PetscInt*,PetscInt*); 395 PETSC_INTERN PetscErrorCode MatCreateMPIMatConcatenateSeqMat_SeqAIJ(MPI_Comm,Mat,PetscInt,MatReuse,Mat*); 396 PETSC_INTERN PetscErrorCode MatCreateMPIMatConcatenateSeqMat_MPIAIJ(MPI_Comm,Mat,PetscInt,MatReuse,Mat*); 397 398 PETSC_INTERN PetscErrorCode MatSetSeqMat_SeqAIJ(Mat,IS,IS,MatStructure,Mat); 399 PETSC_INTERN PetscErrorCode MatDestroySubMatrix_Private(Mat_SubSppt*); 400 PETSC_INTERN PetscErrorCode MatDestroySubMatrix_SeqAIJ(Mat); 401 PETSC_INTERN PetscErrorCode MatDestroySubMatrix_Dummy(Mat); 402 PETSC_INTERN PetscErrorCode MatDestroySubMatrices_Dummy(PetscInt, Mat*[]); 403 PETSC_INTERN PetscErrorCode MatCreateSubMatrix_SeqAIJ(Mat,IS,IS,PetscInt,MatReuse,Mat*); 404 405 PETSC_INTERN PetscErrorCode MatSeqAIJCompactOutExtraColumns_SeqAIJ(Mat,ISLocalToGlobalMapping*); 406 PETSC_INTERN PetscErrorCode MatSetSeqAIJWithArrays_private(MPI_Comm,PetscInt,PetscInt,PetscInt[],PetscInt[],PetscScalar[],MatType,Mat); 407 408 /* 409 PetscSparseDenseMinusDot - The inner kernel of triangular solves and Gauss-Siedel smoothing. \sum_i xv[i] * r[xi[i]] for CSR storage 410 411 Input Parameters: 412 + nnz - the number of entries 413 . r - the array of vector values 414 . xv - the matrix values for the row 415 - xi - the column indices of the nonzeros in the row 416 417 Output Parameter: 418 . sum - negative the sum of results 419 420 PETSc compile flags: 421 + PETSC_KERNEL_USE_UNROLL_4 422 - PETSC_KERNEL_USE_UNROLL_2 423 424 Developer Notes: 425 The macro changes sum but not other parameters 426 427 .seealso: PetscSparseDensePlusDot() 428 429 */ 430 #if defined(PETSC_KERNEL_USE_UNROLL_4) 431 #define PetscSparseDenseMinusDot(sum,r,xv,xi,nnz) { \ 432 if (nnz > 0) { \ 433 PetscInt nnz2=nnz,rem=nnz&0x3; \ 434 switch (rem) { \ 435 case 3: sum -= *xv++ *r[*xi++]; \ 436 case 2: sum -= *xv++ *r[*xi++]; \ 437 case 1: sum -= *xv++ *r[*xi++]; \ 438 nnz2 -= rem;} \ 439 while (nnz2 > 0) { \ 440 sum -= xv[0] * r[xi[0]] + xv[1] * r[xi[1]] + \ 441 xv[2] * r[xi[2]] + xv[3] * r[xi[3]]; \ 442 xv += 4; xi += 4; nnz2 -= 4; \ 443 } \ 444 xv -= nnz; xi -= nnz; \ 445 } \ 446 } 447 448 #elif defined(PETSC_KERNEL_USE_UNROLL_2) 449 #define PetscSparseDenseMinusDot(sum,r,xv,xi,nnz) { \ 450 PetscInt __i,__i1,__i2; \ 451 for (__i=0; __i<nnz-1; __i+=2) {__i1 = xi[__i]; __i2=xi[__i+1]; \ 452 sum -= (xv[__i]*r[__i1] + xv[__i+1]*r[__i2]);} \ 453 if (nnz & 0x1) sum -= xv[__i] * r[xi[__i]];} 454 455 #else 456 #define PetscSparseDenseMinusDot(sum,r,xv,xi,nnz) { \ 457 PetscInt __i; \ 458 for (__i=0; __i<nnz; __i++) sum -= xv[__i] * r[xi[__i]];} 459 #endif 460 461 /* 462 PetscSparseDensePlusDot - The inner kernel of matrix-vector product \sum_i xv[i] * r[xi[i]] for CSR storage 463 464 Input Parameters: 465 + nnz - the number of entries 466 . r - the array of vector values 467 . xv - the matrix values for the row 468 - xi - the column indices of the nonzeros in the row 469 470 Output Parameter: 471 . sum - the sum of results 472 473 PETSc compile flags: 474 + PETSC_KERNEL_USE_UNROLL_4 475 - PETSC_KERNEL_USE_UNROLL_2 476 477 Developer Notes: 478 The macro changes sum but not other parameters 479 480 .seealso: PetscSparseDenseMinusDot() 481 482 */ 483 #if defined(PETSC_KERNEL_USE_UNROLL_4) 484 #define PetscSparseDensePlusDot(sum,r,xv,xi,nnz) { \ 485 if (nnz > 0) { \ 486 PetscInt nnz2=nnz,rem=nnz&0x3; \ 487 switch (rem) { \ 488 case 3: sum += *xv++ *r[*xi++]; \ 489 case 2: sum += *xv++ *r[*xi++]; \ 490 case 1: sum += *xv++ *r[*xi++]; \ 491 nnz2 -= rem;} \ 492 while (nnz2 > 0) { \ 493 sum += xv[0] * r[xi[0]] + xv[1] * r[xi[1]] + \ 494 xv[2] * r[xi[2]] + xv[3] * r[xi[3]]; \ 495 xv += 4; xi += 4; nnz2 -= 4; \ 496 } \ 497 xv -= nnz; xi -= nnz; \ 498 } \ 499 } 500 501 #elif defined(PETSC_KERNEL_USE_UNROLL_2) 502 #define PetscSparseDensePlusDot(sum,r,xv,xi,nnz) { \ 503 PetscInt __i,__i1,__i2; \ 504 for (__i=0; __i<nnz-1; __i+=2) {__i1 = xi[__i]; __i2=xi[__i+1]; \ 505 sum += (xv[__i]*r[__i1] + xv[__i+1]*r[__i2]);} \ 506 if (nnz & 0x1) sum += xv[__i] * r[xi[__i]];} 507 508 #elif defined(PETSC_USE_AVX512_KERNELS) && defined(PETSC_HAVE_IMMINTRIN_H) && defined(__AVX512F__) && defined(PETSC_USE_REAL_DOUBLE) && !defined(PETSC_USE_COMPLEX) && !defined(PETSC_USE_64BIT_INDICES) && !defined(PETSC_SKIP_IMMINTRIN_H_CUDAWORKAROUND) 509 #define PetscSparseDensePlusDot(sum,r,xv,xi,nnz) PetscSparseDensePlusDot_AVX512_Private(&(sum),(r),(xv),(xi),(nnz)) 510 511 #else 512 #define PetscSparseDensePlusDot(sum,r,xv,xi,nnz) { \ 513 PetscInt __i; \ 514 for (__i=0; __i<nnz; __i++) sum += xv[__i] * r[xi[__i]];} 515 #endif 516 517 #if defined(PETSC_USE_AVX512_KERNELS) && defined(PETSC_HAVE_IMMINTRIN_H) && defined(__AVX512F__) && defined(PETSC_USE_REAL_DOUBLE) && !defined(PETSC_USE_COMPLEX) && !defined(PETSC_USE_64BIT_INDICES) && !defined(PETSC_SKIP_IMMINTRIN_H_CUDAWORKAROUND) 518 #include <immintrin.h> 519 #if !defined(_MM_SCALE_8) 520 #define _MM_SCALE_8 8 521 #endif 522 523 static inline void PetscSparseDensePlusDot_AVX512_Private(PetscScalar *sum,const PetscScalar *x,const MatScalar *aa,const PetscInt *aj,PetscInt n) 524 { 525 __m512d vec_x,vec_y,vec_vals; 526 __m256i vec_idx; 527 __mmask8 mask; 528 PetscInt j; 529 530 vec_y = _mm512_setzero_pd(); 531 for (j=0; j<(n>>3); j++) { 532 vec_idx = _mm256_loadu_si256((__m256i const*)aj); 533 vec_vals = _mm512_loadu_pd(aa); 534 vec_x = _mm512_i32gather_pd(vec_idx,x,_MM_SCALE_8); 535 vec_y = _mm512_fmadd_pd(vec_x,vec_vals,vec_y); 536 aj += 8; aa += 8; 537 } 538 /* masked load does not work on KNL, it requires avx512vl */ 539 if ((n&0x07)>2) { 540 mask = (__mmask8)(0xff >> (8-(n&0x07))); 541 vec_idx = _mm256_loadu_si256((__m256i const*)aj); 542 vec_vals = _mm512_loadu_pd(aa); 543 vec_x = _mm512_mask_i32gather_pd(vec_x,mask,vec_idx,x,_MM_SCALE_8); 544 vec_y = _mm512_mask3_fmadd_pd(vec_x,vec_vals,vec_y,mask); 545 } else if ((n&0x07)==2) { 546 *sum += aa[0]*x[aj[0]]; 547 *sum += aa[1]*x[aj[1]]; 548 } else if ((n&0x07)==1) { 549 *sum += aa[0]*x[aj[0]]; 550 } 551 if (n>2) *sum += _mm512_reduce_add_pd(vec_y); 552 /* 553 for (j=0;j<(n&0x07);j++) *sum += aa[j]*x[aj[j]]; 554 */ 555 } 556 #endif 557 558 /* 559 PetscSparseDenseMaxDot - The inner kernel of a modified matrix-vector product \max_i xv[i] * r[xi[i]] for CSR storage 560 561 Input Parameters: 562 + nnz - the number of entries 563 . r - the array of vector values 564 . xv - the matrix values for the row 565 - xi - the column indices of the nonzeros in the row 566 567 Output Parameter: 568 . max - the max of results 569 570 .seealso: PetscSparseDensePlusDot(), PetscSparseDenseMinusDot() 571 572 */ 573 #define PetscSparseDenseMaxDot(max,r,xv,xi,nnz) { \ 574 PetscInt __i; \ 575 for (__i=0; __i<nnz; __i++) max = PetscMax(PetscRealPart(max), PetscRealPart(xv[__i] * r[xi[__i]]));} 576 577 /* 578 Add column indices into table for counting the max nonzeros of merged rows 579 */ 580 #define MatRowMergeMax_SeqAIJ(mat,nrows,ta) { \ 581 PetscInt _j,_row,_nz,*_col; \ 582 if (mat) { \ 583 for (_row=0; _row<nrows; _row++) { \ 584 _nz = mat->i[_row+1] - mat->i[_row]; \ 585 for (_j=0; _j<_nz; _j++) { \ 586 _col = _j + mat->j + mat->i[_row]; \ 587 PetscTableAdd(ta,*_col+1,1,INSERT_VALUES); \ 588 } \ 589 } \ 590 } \ 591 } 592 593 /* 594 Add column indices into table for counting the nonzeros of merged rows 595 */ 596 #define MatMergeRows_SeqAIJ(mat,nrows,rows,ta) { \ 597 PetscInt _j,_row,_nz,*_col,_i; \ 598 for (_i=0; _i<nrows; _i++) {\ 599 _row = rows[_i]; \ 600 _nz = mat->i[_row+1] - mat->i[_row]; \ 601 for (_j=0; _j<_nz; _j++) { \ 602 _col = _j + mat->j + mat->i[_row]; \ 603 PetscTableAdd(ta,*_col+1,1,INSERT_VALUES); \ 604 } \ 605 } \ 606 } 607 608 #endif 609