Actual source code: aij.h


  2: #ifndef __AIJ_H

  5: #include <petsc/private/matimpl.h>
  6: #include <petscctable.h>

  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:   PetscErrorCode (*getcsrandmemtype)(Mat, const PetscInt **, const PetscInt **, PetscScalar **, PetscMemType *);
 17: } Mat_SeqAIJOps;

 19: /*
 20:     Struct header shared by SeqAIJ, SeqBAIJ and SeqSBAIJ matrix formats
 21: */
 22: #define SEQAIJHEADER(datatype) \
 23:   PetscBool         roworiented;  /* if true, row-oriented input, default */ \
 24:   PetscInt          nonew;        /* 1 don't add new nonzeros, -1 generate error on new */ \
 25:   PetscInt          nounused;     /* -1 generate error on unused space */ \
 26:   PetscBool         singlemalloc; /* if true a, i, and j have been obtained with one big malloc */ \
 27:   PetscInt          maxnz;        /* allocated nonzeros */ \
 28:   PetscInt         *imax;         /* maximum space allocated for each row */ \
 29:   PetscInt         *ilen;         /* actual length of each row */ \
 30:   PetscInt         *ipre;         /* space preallocated for each row by user */ \
 31:   PetscBool         free_imax_ilen; \
 32:   PetscInt          reallocs;           /* number of mallocs done during MatSetValues() \
 33:                                         as more values are set than were prealloced */ \
 34:   PetscInt          rmax;               /* max nonzeros in any row */ \
 35:   PetscBool         keepnonzeropattern; /* keeps matrix structure same in calls to MatZeroRows()*/ \
 36:   PetscBool         ignorezeroentries; \
 37:   PetscBool         free_ij;       /* free the column indices j and row offsets i when the matrix is destroyed */ \
 38:   PetscBool         free_a;        /* free the numerical values when matrix is destroy */ \
 39:   Mat_CompressedRow compressedrow; /* use compressed row format */ \
 40:   PetscInt          nz;            /* nonzeros */ \
 41:   PetscInt         *i;             /* pointer to beginning of each row */ \
 42:   PetscInt         *j;             /* column values: j + i[k] - 1 is start of row k */ \
 43:   PetscInt         *diag;          /* pointers to diagonal elements */ \
 44:   PetscInt          nonzerorowcnt; /* how many rows have nonzero entries */ \
 45:   PetscBool         free_diag; \
 46:   datatype         *a;              /* nonzero elements */ \
 47:   PetscScalar      *solve_work;     /* work space used in MatSolve */ \
 48:   IS                row, col, icol; /* index sets, used for reorderings */ \
 49:   PetscBool         pivotinblocks;  /* pivot inside factorization of each diagonal block */ \
 50:   Mat               parent;         /* set if this matrix was formed with MatDuplicate(...,MAT_SHARE_NONZERO_PATTERN,....); \
 51:                                          means that this shares some data structures with the parent including diag, ilen, imax, i, j */ \
 52:   Mat_SubSppt      *submatis1;      /* used by MatCreateSubMatrices_MPIXAIJ_Local */ \
 53:   Mat_SeqAIJOps     ops[1]          /* operations for SeqAIJ and its subclasses */

 55: typedef struct {
 56:   MatTransposeColoring matcoloring;
 57:   Mat                  Bt_den;  /* dense matrix of B^T */
 58:   Mat                  ABt_den; /* dense matrix of A*B^T */
 59:   PetscBool            usecoloring;
 60: } Mat_MatMatTransMult;

 62: typedef struct { /* used by MatTransposeMatMult() */
 63:   Mat At;        /* transpose of the first matrix */
 64:   Mat mA;        /* maij matrix of A */
 65:   Vec bt, ct;    /* vectors to hold locally transposed arrays of B and C */
 66:   /* used by PtAP */
 67:   void *data;
 68:   PetscErrorCode (*destroy)(void *);
 69: } Mat_MatTransMatMult;

 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;

 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;

 87: typedef struct {
 88:   Mat BC; /* temp matrix for storing B*C */
 89: } Mat_MatMatMatMult;

 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: */

 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 */

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;

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 **);

125: typedef struct {
126:   SEQAIJHEADER(MatScalar);
127:   Mat_SeqAIJ_Inode inode;
128:   MatScalar       *saved_values; /* location for stashing nonzero values of matrix */

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 */

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;

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:     PetscFree3(*a, *j, *i);
152:   } else {
153:     if (A->free_a) PetscFree(*a);
154:     if (A->free_ij) PetscFree(*j);
155:     if (A->free_ij) 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: \
171:     /* malloc new storage space */ \
172:     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:     PetscArraycpy(new_j, AJ, AI[ROW] + NROW); \
178:     len = (new_nz - CHUNKSIZE - AI[ROW] - NROW); \
179:     PetscArraycpy(new_j + AI[ROW] + NROW + CHUNKSIZE, AJ + AI[ROW] + NROW, len); \
180:     PetscArraycpy(new_a, AA, BS2 *(AI[ROW] + NROW)); \
181:     PetscArrayzero(new_a + BS2 * (AI[ROW] + NROW), BS2 * CHUNKSIZE); \
182:     PetscArraycpy(new_a + BS2 * (AI[ROW] + NROW + CHUNKSIZE), AA + BS2 * (AI[ROW] + NROW), BS2 * len); \
183:     /* free up old matrix storage */ \
184:     MatSeqXAIJFreeAIJ(A, &Ain->a, &Ain->j, &Ain->i); \
185:     AA     = new_a; \
186:     Ain->a = (MatScalar *)new_a; \
187:     AI = Ain->i = new_i; \
188:     AJ = Ain->j       = new_j; \
189:     Ain->singlemalloc = PETSC_TRUE; \
190: \
191:     RP   = AJ + AI[ROW]; \
192:     AP   = AA + BS2 * AI[ROW]; \
193:     RMAX = AIMAX[ROW] = AIMAX[ROW] + CHUNKSIZE; \
194:     Ain->maxnz += BS2 * CHUNKSIZE; \
195:     Ain->reallocs++; \
196:   }

198: #define MatSeqXAIJReallocateAIJ_structure_only(Amat, AM, BS2, NROW, ROW, COL, RMAX, AI, AJ, RP, AIMAX, NONEW, datatype) \
199:   if (NROW >= RMAX) { \
200:     Mat_SeqAIJ *Ain = (Mat_SeqAIJ *)Amat->data; \
201:     /* there is no extra room in row, therefore enlarge */ \
202:     PetscInt CHUNKSIZE = 15, new_nz = AI[AM] + CHUNKSIZE, len, *new_i = NULL, *new_j = NULL; \
203: \
205:     /* malloc new storage space */ \
206:     PetscMalloc1(new_nz, &new_j); \
207:     PetscMalloc1(AM + 1, &new_i); \
208: \
209:     /* copy over old data into new slots */ \
210:     for (ii = 0; ii < ROW + 1; ii++) new_i[ii] = AI[ii]; \
211:     for (ii = ROW + 1; ii < AM + 1; ii++) new_i[ii] = AI[ii] + CHUNKSIZE; \
212:     PetscArraycpy(new_j, AJ, AI[ROW] + NROW); \
213:     len = (new_nz - CHUNKSIZE - AI[ROW] - NROW); \
214:     PetscArraycpy(new_j + AI[ROW] + NROW + CHUNKSIZE, AJ + AI[ROW] + NROW, len); \
215: \
216:     /* free up old matrix storage */ \
217:     MatSeqXAIJFreeAIJ(A, &Ain->a, &Ain->j, &Ain->i); \
218:     Ain->a = NULL; \
219:     AI = Ain->i = new_i; \
220:     AJ = Ain->j       = new_j; \
221:     Ain->singlemalloc = PETSC_FALSE; \
222:     Ain->free_a       = PETSC_FALSE; \
223: \
224:     RP   = AJ + AI[ROW]; \
225:     RMAX = AIMAX[ROW] = AIMAX[ROW] + CHUNKSIZE; \
226:     Ain->maxnz += BS2 * CHUNKSIZE; \
227:     Ain->reallocs++; \
228:   }

230: PETSC_INTERN PetscErrorCode MatSeqAIJSetPreallocation_SeqAIJ(Mat, PetscInt, const PetscInt *);
231: PETSC_INTERN PetscErrorCode MatSetPreallocationCOO_SeqAIJ(Mat, PetscCount, PetscInt[], PetscInt[]);
232: PETSC_INTERN PetscErrorCode MatResetPreallocationCOO_SeqAIJ(Mat);

234: PETSC_INTERN PetscErrorCode MatILUFactorSymbolic_SeqAIJ_inplace(Mat, Mat, IS, IS, const MatFactorInfo *);
235: PETSC_INTERN PetscErrorCode MatILUFactorSymbolic_SeqAIJ(Mat, Mat, IS, IS, const MatFactorInfo *);
236: PETSC_INTERN PetscErrorCode MatILUFactorSymbolic_SeqAIJ_ilu0(Mat, Mat, IS, IS, const MatFactorInfo *);

238: PETSC_INTERN PetscErrorCode MatICCFactorSymbolic_SeqAIJ_inplace(Mat, Mat, IS, const MatFactorInfo *);
239: PETSC_INTERN PetscErrorCode MatICCFactorSymbolic_SeqAIJ(Mat, Mat, IS, const MatFactorInfo *);
240: PETSC_INTERN PetscErrorCode MatCholeskyFactorSymbolic_SeqAIJ_inplace(Mat, Mat, IS, const MatFactorInfo *);
241: PETSC_INTERN PetscErrorCode MatCholeskyFactorSymbolic_SeqAIJ(Mat, Mat, IS, const MatFactorInfo *);
242: PETSC_INTERN PetscErrorCode MatCholeskyFactorNumeric_SeqAIJ_inplace(Mat, Mat, const MatFactorInfo *);
243: PETSC_INTERN PetscErrorCode MatCholeskyFactorNumeric_SeqAIJ(Mat, Mat, const MatFactorInfo *);
244: PETSC_INTERN PetscErrorCode MatDuplicate_SeqAIJ(Mat, MatDuplicateOption, Mat *);
245: PETSC_INTERN PetscErrorCode MatCopy_SeqAIJ(Mat, Mat, MatStructure);
246: PETSC_INTERN PetscErrorCode MatMissingDiagonal_SeqAIJ(Mat, PetscBool *, PetscInt *);
247: PETSC_INTERN PetscErrorCode MatMarkDiagonal_SeqAIJ(Mat);
248: PETSC_INTERN PetscErrorCode MatFindZeroDiagonals_SeqAIJ_Private(Mat, PetscInt *, PetscInt **);

250: PETSC_INTERN PetscErrorCode MatMult_SeqAIJ(Mat, Vec, Vec);
251: PETSC_INTERN PetscErrorCode MatMult_SeqAIJ_Inode(Mat, Vec, Vec);
252: PETSC_INTERN PetscErrorCode MatMultAdd_SeqAIJ(Mat, Vec, Vec, Vec);
253: PETSC_INTERN PetscErrorCode MatMultAdd_SeqAIJ_Inode(Mat, Vec, Vec, Vec);
254: PETSC_INTERN PetscErrorCode MatMultTranspose_SeqAIJ(Mat, Vec, Vec);
255: PETSC_INTERN PetscErrorCode MatMultTransposeAdd_SeqAIJ(Mat, Vec, Vec, Vec);
256: PETSC_INTERN PetscErrorCode MatSOR_SeqAIJ(Mat, Vec, PetscReal, MatSORType, PetscReal, PetscInt, PetscInt, Vec);
257: PETSC_INTERN PetscErrorCode MatSOR_SeqAIJ_Inode(Mat, Vec, PetscReal, MatSORType, PetscReal, PetscInt, PetscInt, Vec);

259: PETSC_INTERN PetscErrorCode MatSetOption_SeqAIJ(Mat, MatOption, PetscBool);

261: PETSC_INTERN PetscErrorCode MatGetSymbolicTranspose_SeqAIJ(Mat, PetscInt *[], PetscInt *[]);
262: PETSC_INTERN PetscErrorCode MatRestoreSymbolicTranspose_SeqAIJ(Mat, PetscInt *[], PetscInt *[]);
263: PETSC_INTERN PetscErrorCode MatGetSymbolicTransposeReduced_SeqAIJ(Mat, PetscInt, PetscInt, PetscInt *[], PetscInt *[]);
264: PETSC_INTERN PetscErrorCode MatTransposeSymbolic_SeqAIJ(Mat, Mat *);
265: PETSC_INTERN PetscErrorCode MatTranspose_SeqAIJ(Mat, MatReuse, Mat *);

267: PETSC_INTERN PetscErrorCode MatToSymmetricIJ_SeqAIJ(PetscInt, PetscInt *, PetscInt *, PetscBool, PetscInt, PetscInt, PetscInt **, PetscInt **);
268: PETSC_INTERN PetscErrorCode MatLUFactorSymbolic_SeqAIJ_inplace(Mat, Mat, IS, IS, const MatFactorInfo *);
269: PETSC_INTERN PetscErrorCode MatLUFactorSymbolic_SeqAIJ(Mat, Mat, IS, IS, const MatFactorInfo *);
270: PETSC_INTERN PetscErrorCode MatLUFactorNumeric_SeqAIJ_inplace(Mat, Mat, const MatFactorInfo *);
271: PETSC_INTERN PetscErrorCode MatLUFactorNumeric_SeqAIJ(Mat, Mat, const MatFactorInfo *);
272: PETSC_INTERN PetscErrorCode MatLUFactorNumeric_SeqAIJ_InplaceWithPerm(Mat, Mat, const MatFactorInfo *);
273: PETSC_INTERN PetscErrorCode MatLUFactor_SeqAIJ(Mat, IS, IS, const MatFactorInfo *);
274: PETSC_INTERN PetscErrorCode MatSolve_SeqAIJ_inplace(Mat, Vec, Vec);
275: PETSC_INTERN PetscErrorCode MatSolve_SeqAIJ(Mat, Vec, Vec);
276: PETSC_INTERN PetscErrorCode MatSolve_SeqAIJ_Inode_inplace(Mat, Vec, Vec);
277: PETSC_INTERN PetscErrorCode MatSolve_SeqAIJ_Inode(Mat, Vec, Vec);
278: PETSC_INTERN PetscErrorCode MatSolve_SeqAIJ_NaturalOrdering_inplace(Mat, Vec, Vec);
279: PETSC_INTERN PetscErrorCode MatSolve_SeqAIJ_NaturalOrdering(Mat, Vec, Vec);
280: PETSC_INTERN PetscErrorCode MatSolve_SeqAIJ_InplaceWithPerm(Mat, Vec, Vec);
281: PETSC_INTERN PetscErrorCode MatSolveAdd_SeqAIJ_inplace(Mat, Vec, Vec, Vec);
282: PETSC_INTERN PetscErrorCode MatSolveAdd_SeqAIJ(Mat, Vec, Vec, Vec);
283: PETSC_INTERN PetscErrorCode MatSolveTranspose_SeqAIJ_inplace(Mat, Vec, Vec);
284: PETSC_INTERN PetscErrorCode MatSolveTranspose_SeqAIJ(Mat, Vec, Vec);
285: PETSC_INTERN PetscErrorCode MatSolveTransposeAdd_SeqAIJ_inplace(Mat, Vec, Vec, Vec);
286: PETSC_INTERN PetscErrorCode MatSolveTransposeAdd_SeqAIJ(Mat, Vec, Vec, Vec);
287: PETSC_INTERN PetscErrorCode MatMatSolve_SeqAIJ_inplace(Mat, Mat, Mat);
288: PETSC_INTERN PetscErrorCode MatMatSolve_SeqAIJ(Mat, Mat, Mat);
289: PETSC_INTERN PetscErrorCode MatEqual_SeqAIJ(Mat, Mat, PetscBool *);
290: PETSC_INTERN PetscErrorCode MatFDColoringCreate_SeqXAIJ(Mat, ISColoring, MatFDColoring);
291: PETSC_INTERN PetscErrorCode MatFDColoringSetUp_SeqXAIJ(Mat, ISColoring, MatFDColoring);
292: PETSC_INTERN PetscErrorCode MatFDColoringSetUpBlocked_AIJ_Private(Mat, MatFDColoring, PetscInt);
293: PETSC_INTERN PetscErrorCode MatLoad_AIJ_HDF5(Mat, PetscViewer);
294: PETSC_INTERN PetscErrorCode MatLoad_SeqAIJ_Binary(Mat, PetscViewer);
295: PETSC_INTERN PetscErrorCode MatLoad_SeqAIJ(Mat, PetscViewer);
296: PETSC_INTERN PetscErrorCode RegisterApplyPtAPRoutines_Private(Mat);

298: #if defined(PETSC_HAVE_HYPRE)
299: PETSC_INTERN PetscErrorCode MatProductSetFromOptions_Transpose_AIJ_AIJ(Mat);
300: #endif
301: PETSC_INTERN PetscErrorCode MatProductSetFromOptions_SeqAIJ(Mat);

303: PETSC_INTERN PetscErrorCode MatProductSymbolic_SeqAIJ_SeqAIJ(Mat);
304: PETSC_INTERN PetscErrorCode MatProductSymbolic_PtAP_SeqAIJ_SeqAIJ(Mat);
305: PETSC_INTERN PetscErrorCode MatProductSymbolic_RARt_SeqAIJ_SeqAIJ(Mat);

307: PETSC_INTERN PetscErrorCode MatMatMultSymbolic_SeqAIJ_SeqAIJ(Mat, Mat, PetscReal, Mat);
308: PETSC_INTERN PetscErrorCode MatMatMultSymbolic_SeqAIJ_SeqAIJ_Sorted(Mat, Mat, PetscReal, Mat);
309: PETSC_INTERN PetscErrorCode MatMatMultSymbolic_SeqDense_SeqAIJ(Mat, Mat, PetscReal, Mat);
310: PETSC_INTERN PetscErrorCode MatMatMultSymbolic_SeqAIJ_SeqAIJ_Scalable(Mat, Mat, PetscReal, Mat);
311: PETSC_INTERN PetscErrorCode MatMatMultSymbolic_SeqAIJ_SeqAIJ_Scalable_fast(Mat, Mat, PetscReal, Mat);
312: PETSC_INTERN PetscErrorCode MatMatMultSymbolic_SeqAIJ_SeqAIJ_Heap(Mat, Mat, PetscReal, Mat);
313: PETSC_INTERN PetscErrorCode MatMatMultSymbolic_SeqAIJ_SeqAIJ_BTHeap(Mat, Mat, PetscReal, Mat);
314: PETSC_INTERN PetscErrorCode MatMatMultSymbolic_SeqAIJ_SeqAIJ_RowMerge(Mat, Mat, PetscReal, Mat);
315: PETSC_INTERN PetscErrorCode MatMatMultSymbolic_SeqAIJ_SeqAIJ_LLCondensed(Mat, Mat, PetscReal, Mat);
316: #if defined(PETSC_HAVE_HYPRE)
317: PETSC_INTERN PetscErrorCode MatMatMultSymbolic_AIJ_AIJ_wHYPRE(Mat, Mat, PetscReal, Mat);
318: #endif

320: PETSC_INTERN PetscErrorCode MatMatMultNumeric_SeqAIJ_SeqAIJ(Mat, Mat, Mat);
321: PETSC_INTERN PetscErrorCode MatMatMultNumeric_SeqAIJ_SeqAIJ_Sorted(Mat, Mat, Mat);

323: PETSC_INTERN PetscErrorCode MatMatMultNumeric_SeqDense_SeqAIJ(Mat, Mat, Mat);
324: PETSC_INTERN PetscErrorCode MatMatMultNumeric_SeqAIJ_SeqAIJ_Scalable(Mat, Mat, Mat);

326: PETSC_INTERN PetscErrorCode MatPtAPSymbolic_SeqAIJ_SeqAIJ_SparseAxpy(Mat, Mat, PetscReal, Mat);
327: PETSC_INTERN PetscErrorCode MatPtAPNumeric_SeqAIJ_SeqAIJ(Mat, Mat, Mat);
328: PETSC_INTERN PetscErrorCode MatPtAPNumeric_SeqAIJ_SeqAIJ_SparseAxpy(Mat, Mat, Mat);

330: PETSC_INTERN PetscErrorCode MatRARtSymbolic_SeqAIJ_SeqAIJ(Mat, Mat, PetscReal, Mat);
331: PETSC_INTERN PetscErrorCode MatRARtSymbolic_SeqAIJ_SeqAIJ_matmattransposemult(Mat, Mat, PetscReal, Mat);
332: PETSC_INTERN PetscErrorCode MatRARtSymbolic_SeqAIJ_SeqAIJ_colorrart(Mat, Mat, PetscReal, Mat);
333: PETSC_INTERN PetscErrorCode MatRARtNumeric_SeqAIJ_SeqAIJ(Mat, Mat, Mat);
334: PETSC_INTERN PetscErrorCode MatRARtNumeric_SeqAIJ_SeqAIJ_matmattransposemult(Mat, Mat, Mat);
335: PETSC_INTERN PetscErrorCode MatRARtNumeric_SeqAIJ_SeqAIJ_colorrart(Mat, Mat, Mat);

337: PETSC_INTERN PetscErrorCode MatTransposeMatMultSymbolic_SeqAIJ_SeqAIJ(Mat, Mat, PetscReal, Mat);
338: PETSC_INTERN PetscErrorCode MatTransposeMatMultNumeric_SeqAIJ_SeqAIJ(Mat, Mat, Mat);
339: PETSC_INTERN PetscErrorCode MatDestroy_SeqAIJ_MatTransMatMult(void *);

341: PETSC_INTERN PetscErrorCode MatMatTransposeMultSymbolic_SeqAIJ_SeqAIJ(Mat, Mat, PetscReal, Mat);
342: PETSC_INTERN PetscErrorCode MatMatTransposeMultNumeric_SeqAIJ_SeqAIJ(Mat, Mat, Mat);
343: PETSC_INTERN PetscErrorCode MatTransposeColoringCreate_SeqAIJ(Mat, ISColoring, MatTransposeColoring);
344: PETSC_INTERN PetscErrorCode MatTransColoringApplySpToDen_SeqAIJ(MatTransposeColoring, Mat, Mat);
345: PETSC_INTERN PetscErrorCode MatTransColoringApplyDenToSp_SeqAIJ(MatTransposeColoring, Mat, Mat);

347: PETSC_INTERN PetscErrorCode MatMatMatMultSymbolic_SeqAIJ_SeqAIJ_SeqAIJ(Mat, Mat, Mat, PetscReal, Mat);
348: PETSC_INTERN PetscErrorCode MatMatMatMultNumeric_SeqAIJ_SeqAIJ_SeqAIJ(Mat, Mat, Mat, Mat);

350: PETSC_INTERN PetscErrorCode MatSetRandomSkipColumnRange_SeqAIJ_Private(Mat, PetscInt, PetscInt, PetscRandom);
351: PETSC_INTERN PetscErrorCode MatSetValues_SeqAIJ(Mat, PetscInt, const PetscInt[], PetscInt, const PetscInt[], const PetscScalar[], InsertMode);
352: PETSC_INTERN PetscErrorCode MatGetRow_SeqAIJ(Mat, PetscInt, PetscInt *, PetscInt **, PetscScalar **);
353: PETSC_INTERN PetscErrorCode MatRestoreRow_SeqAIJ(Mat, PetscInt, PetscInt *, PetscInt **, PetscScalar **);
354: PETSC_INTERN PetscErrorCode MatScale_SeqAIJ(Mat, PetscScalar);
355: PETSC_INTERN PetscErrorCode MatDiagonalScale_SeqAIJ(Mat, Vec, Vec);
356: PETSC_INTERN PetscErrorCode MatDiagonalSet_SeqAIJ(Mat, Vec, InsertMode);
357: PETSC_INTERN PetscErrorCode MatAXPY_SeqAIJ(Mat, PetscScalar, Mat, MatStructure);
358: PETSC_INTERN PetscErrorCode MatGetRowIJ_SeqAIJ(Mat, PetscInt, PetscBool, PetscBool, PetscInt *, const PetscInt *[], const PetscInt *[], PetscBool *);
359: PETSC_INTERN PetscErrorCode MatRestoreRowIJ_SeqAIJ(Mat, PetscInt, PetscBool, PetscBool, PetscInt *, const PetscInt *[], const PetscInt *[], PetscBool *);
360: PETSC_INTERN PetscErrorCode MatGetColumnIJ_SeqAIJ(Mat, PetscInt, PetscBool, PetscBool, PetscInt *, const PetscInt *[], const PetscInt *[], PetscBool *);
361: PETSC_INTERN PetscErrorCode MatRestoreColumnIJ_SeqAIJ(Mat, PetscInt, PetscBool, PetscBool, PetscInt *, const PetscInt *[], const PetscInt *[], PetscBool *);
362: PETSC_INTERN PetscErrorCode MatGetColumnIJ_SeqAIJ_Color(Mat, PetscInt, PetscBool, PetscBool, PetscInt *, const PetscInt *[], const PetscInt *[], PetscInt *[], PetscBool *);
363: PETSC_INTERN PetscErrorCode MatRestoreColumnIJ_SeqAIJ_Color(Mat, PetscInt, PetscBool, PetscBool, PetscInt *, const PetscInt *[], const PetscInt *[], PetscInt *[], PetscBool *);
364: PETSC_INTERN PetscErrorCode MatDestroy_SeqAIJ(Mat);
365: PETSC_INTERN PetscErrorCode MatSetUp_SeqAIJ(Mat);
366: PETSC_INTERN PetscErrorCode MatView_SeqAIJ(Mat, PetscViewer);

368: PETSC_INTERN PetscErrorCode MatSeqAIJInvalidateDiagonal(Mat);
369: PETSC_INTERN PetscErrorCode MatSeqAIJInvalidateDiagonal_Inode(Mat);
370: PETSC_INTERN PetscErrorCode MatSeqAIJCheckInode(Mat);
371: PETSC_INTERN PetscErrorCode MatSeqAIJCheckInode_FactorLU(Mat);

373: PETSC_INTERN PetscErrorCode MatAXPYGetPreallocation_SeqAIJ(Mat, Mat, PetscInt *);

375: #if defined(PETSC_HAVE_MATLAB)
376: PETSC_EXTERN PetscErrorCode MatlabEnginePut_SeqAIJ(PetscObject, void *);
377: PETSC_EXTERN PetscErrorCode MatlabEngineGet_SeqAIJ(PetscObject, void *);
378: #endif
379: PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqSBAIJ(Mat, MatType, MatReuse, Mat *);
380: PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqBAIJ(Mat, MatType, MatReuse, Mat *);
381: PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqDense(Mat, MatType, MatReuse, Mat *);
382: PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqAIJCRL(Mat, MatType, MatReuse, Mat *);
383: PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_Elemental(Mat, MatType, MatReuse, Mat *);
384: #if defined(PETSC_HAVE_SCALAPACK)
385: PETSC_INTERN PetscErrorCode MatConvert_AIJ_ScaLAPACK(Mat, MatType, MatReuse, Mat *);
386: #endif
387: PETSC_INTERN PetscErrorCode MatConvert_AIJ_HYPRE(Mat, MatType, MatReuse, Mat *);
388: PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqAIJPERM(Mat, MatType, MatReuse, Mat *);
389: PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqAIJSELL(Mat, MatType, MatReuse, Mat *);
390: PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqAIJMKL(Mat, MatType, MatReuse, Mat *);
391: PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqAIJViennaCL(Mat, MatType, MatReuse, Mat *);
392: PETSC_INTERN PetscErrorCode MatReorderForNonzeroDiagonal_SeqAIJ(Mat, PetscReal, IS, IS);
393: PETSC_INTERN PetscErrorCode MatRARt_SeqAIJ_SeqAIJ(Mat, Mat, MatReuse, PetscReal, Mat *);
394: PETSC_EXTERN PetscErrorCode MatCreate_SeqAIJ(Mat);
395: PETSC_INTERN PetscErrorCode MatAssemblyEnd_SeqAIJ(Mat, MatAssemblyType);
396: PETSC_EXTERN PetscErrorCode MatZeroEntries_SeqAIJ(Mat);

398: PETSC_INTERN PetscErrorCode MatAXPYGetPreallocation_SeqX_private(PetscInt, const PetscInt *, const PetscInt *, const PetscInt *, const PetscInt *, PetscInt *);
399: PETSC_INTERN PetscErrorCode MatCreateMPIMatConcatenateSeqMat_SeqAIJ(MPI_Comm, Mat, PetscInt, MatReuse, Mat *);
400: PETSC_INTERN PetscErrorCode MatCreateMPIMatConcatenateSeqMat_MPIAIJ(MPI_Comm, Mat, PetscInt, MatReuse, Mat *);

402: PETSC_INTERN PetscErrorCode MatSetSeqMat_SeqAIJ(Mat, IS, IS, MatStructure, Mat);
403: PETSC_INTERN PetscErrorCode MatDestroySubMatrix_Private(Mat_SubSppt *);
404: PETSC_INTERN PetscErrorCode MatDestroySubMatrix_SeqAIJ(Mat);
405: PETSC_INTERN PetscErrorCode MatDestroySubMatrix_Dummy(Mat);
406: PETSC_INTERN PetscErrorCode MatDestroySubMatrices_Dummy(PetscInt, Mat *[]);
407: PETSC_INTERN PetscErrorCode MatCreateSubMatrix_SeqAIJ(Mat, IS, IS, PetscInt, MatReuse, Mat *);

409: PETSC_INTERN PetscErrorCode MatSeqAIJCompactOutExtraColumns_SeqAIJ(Mat, ISLocalToGlobalMapping *);
410: PETSC_INTERN PetscErrorCode MatSetSeqAIJWithArrays_private(MPI_Comm, PetscInt, PetscInt, PetscInt[], PetscInt[], PetscScalar[], MatType, Mat);

412: /*
413:     PetscSparseDenseMinusDot - The inner kernel of triangular solves and Gauss-Siedel smoothing. \sum_i xv[i] * r[xi[i]] for CSR storage

415:   Input Parameters:
416: +  nnz - the number of entries
417: .  r - the array of vector values
418: .  xv - the matrix values for the row
419: -  xi - the column indices of the nonzeros in the row

421:   Output Parameter:
422: .  sum - negative the sum of results

424:   PETSc compile flags:
425: +   PETSC_KERNEL_USE_UNROLL_4
426: -   PETSC_KERNEL_USE_UNROLL_2

428:   Developer Note:
429:     The macro changes sum but not other parameters

431: .seealso: `PetscSparseDensePlusDot()`
432: */
433: #if defined(PETSC_KERNEL_USE_UNROLL_4)
434:   #define PetscSparseDenseMinusDot(sum, r, xv, xi, nnz) \
435:     { \
436:       if (nnz > 0) { \
437:         PetscInt nnz2 = nnz, rem = nnz & 0x3; \
438:         switch (rem) { \
439:         case 3: \
440:           sum -= *xv++ * r[*xi++]; \
441:         case 2: \
442:           sum -= *xv++ * r[*xi++]; \
443:         case 1: \
444:           sum -= *xv++ * r[*xi++]; \
445:           nnz2 -= rem; \
446:         } \
447:         while (nnz2 > 0) { \
448:           sum -= xv[0] * r[xi[0]] + xv[1] * r[xi[1]] + xv[2] * r[xi[2]] + xv[3] * r[xi[3]]; \
449:           xv += 4; \
450:           xi += 4; \
451:           nnz2 -= 4; \
452:         } \
453:         xv -= nnz; \
454:         xi -= nnz; \
455:       } \
456:     }

458: #elif defined(PETSC_KERNEL_USE_UNROLL_2)
459:   #define PetscSparseDenseMinusDot(sum, r, xv, xi, nnz) \
460:     { \
461:       PetscInt __i, __i1, __i2; \
462:       for (__i = 0; __i < nnz - 1; __i += 2) { \
463:         __i1 = xi[__i]; \
464:         __i2 = xi[__i + 1]; \
465:         sum -= (xv[__i] * r[__i1] + xv[__i + 1] * r[__i2]); \
466:       } \
467:       if (nnz & 0x1) sum -= xv[__i] * r[xi[__i]]; \
468:     }

470: #else
471:   #define PetscSparseDenseMinusDot(sum, r, xv, xi, nnz) \
472:     { \
473:       PetscInt __i; \
474:       for (__i = 0; __i < nnz; __i++) sum -= xv[__i] * r[xi[__i]]; \
475:     }
476: #endif

478: /*
479:     PetscSparseDensePlusDot - The inner kernel of matrix-vector product \sum_i xv[i] * r[xi[i]] for CSR storage

481:   Input Parameters:
482: +  nnz - the number of entries
483: .  r - the array of vector values
484: .  xv - the matrix values for the row
485: -  xi - the column indices of the nonzeros in the row

487:   Output Parameter:
488: .  sum - the sum of results

490:   PETSc compile flags:
491: +   PETSC_KERNEL_USE_UNROLL_4
492: -   PETSC_KERNEL_USE_UNROLL_2

494:   Developer Note:
495:     The macro changes sum but not other parameters

497: .seealso: `PetscSparseDenseMinusDot()`
498: */
499: #if defined(PETSC_KERNEL_USE_UNROLL_4)
500:   #define PetscSparseDensePlusDot(sum, r, xv, xi, nnz) \
501:     { \
502:       if (nnz > 0) { \
503:         PetscInt nnz2 = nnz, rem = nnz & 0x3; \
504:         switch (rem) { \
505:         case 3: \
506:           sum += *xv++ * r[*xi++]; \
507:         case 2: \
508:           sum += *xv++ * r[*xi++]; \
509:         case 1: \
510:           sum += *xv++ * r[*xi++]; \
511:           nnz2 -= rem; \
512:         } \
513:         while (nnz2 > 0) { \
514:           sum += xv[0] * r[xi[0]] + xv[1] * r[xi[1]] + xv[2] * r[xi[2]] + xv[3] * r[xi[3]]; \
515:           xv += 4; \
516:           xi += 4; \
517:           nnz2 -= 4; \
518:         } \
519:         xv -= nnz; \
520:         xi -= nnz; \
521:       } \
522:     }

524: #elif defined(PETSC_KERNEL_USE_UNROLL_2)
525:   #define PetscSparseDensePlusDot(sum, r, xv, xi, nnz) \
526:     { \
527:       PetscInt __i, __i1, __i2; \
528:       for (__i = 0; __i < nnz - 1; __i += 2) { \
529:         __i1 = xi[__i]; \
530:         __i2 = xi[__i + 1]; \
531:         sum += (xv[__i] * r[__i1] + xv[__i + 1] * r[__i2]); \
532:       } \
533:       if (nnz & 0x1) sum += xv[__i] * r[xi[__i]]; \
534:     }

536: #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)
537:   #define PetscSparseDensePlusDot(sum, r, xv, xi, nnz) PetscSparseDensePlusDot_AVX512_Private(&(sum), (r), (xv), (xi), (nnz))

539: #else
540:   #define PetscSparseDensePlusDot(sum, r, xv, xi, nnz) \
541:     { \
542:       PetscInt __i; \
543:       for (__i = 0; __i < nnz; __i++) sum += xv[__i] * r[xi[__i]]; \
544:     }
545: #endif

547: #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)
548:   #include <immintrin.h>
549:   #if !defined(_MM_SCALE_8)
550:     #define _MM_SCALE_8 8
551:   #endif

553: static inline void PetscSparseDensePlusDot_AVX512_Private(PetscScalar *sum, const PetscScalar *x, const MatScalar *aa, const PetscInt *aj, PetscInt n)
554: {
555:   __m512d  vec_x, vec_y, vec_vals;
556:   __m256i  vec_idx;
557:   PetscInt j;

559:   vec_y = _mm512_setzero_pd();
560:   for (j = 0; j < (n >> 3); j++) {
561:     vec_idx  = _mm256_loadu_si256((__m256i const *)aj);
562:     vec_vals = _mm512_loadu_pd(aa);
563:     vec_x    = _mm512_i32gather_pd(vec_idx, x, _MM_SCALE_8);
564:     vec_y    = _mm512_fmadd_pd(vec_x, vec_vals, vec_y);
565:     aj += 8;
566:     aa += 8;
567:   }
568:   #if defined(__AVX512VL__)
569:   /* masked load requires avx512vl, which is not supported by KNL */
570:   if (n & 0x07) {
571:     __mmask8 mask;
572:     mask     = (__mmask8)(0xff >> (8 - (n & 0x07)));
573:     vec_idx  = _mm256_mask_loadu_epi32(vec_idx, mask, aj);
574:     vec_vals = _mm512_mask_loadu_pd(vec_vals, mask, aa);
575:     vec_x    = _mm512_mask_i32gather_pd(vec_x, mask, vec_idx, x, _MM_SCALE_8);
576:     vec_y    = _mm512_mask3_fmadd_pd(vec_x, vec_vals, vec_y, mask);
577:   }
578:   *sum += _mm512_reduce_add_pd(vec_y);
579:   #else
580:   *sum += _mm512_reduce_add_pd(vec_y);
581:   for (j = 0; j < (n & 0x07); j++) *sum += aa[j] * x[aj[j]];
582:   #endif
583: }
584: #endif

586: /*
587:     PetscSparseDenseMaxDot - The inner kernel of a modified matrix-vector product \max_i xv[i] * r[xi[i]] for CSR storage

589:   Input Parameters:
590: +  nnz - the number of entries
591: .  r - the array of vector values
592: .  xv - the matrix values for the row
593: -  xi - the column indices of the nonzeros in the row

595:   Output Parameter:
596: .  max - the max of results

598: .seealso: `PetscSparseDensePlusDot()`, `PetscSparseDenseMinusDot()`
599: */
600: #define PetscSparseDenseMaxDot(max, r, xv, xi, nnz) \
601:   { \
602:     PetscInt __i; \
603:     for (__i = 0; __i < nnz; __i++) max = PetscMax(PetscRealPart(max), PetscRealPart(xv[__i] * r[xi[__i]])); \
604:   }

606: /*
607:  Add column indices into table for counting the max nonzeros of merged rows
608:  */
609: #define MatRowMergeMax_SeqAIJ(mat, nrows, ta) \
610:   { \
611:     PetscInt _j, _row, _nz, *_col; \
612:     if (mat) { \
613:       for (_row = 0; _row < nrows; _row++) { \
614:         _nz = mat->i[_row + 1] - mat->i[_row]; \
615:         for (_j = 0; _j < _nz; _j++) { \
616:           _col = _j + mat->j + mat->i[_row]; \
617:           PetscTableAdd(ta, *_col + 1, 1, INSERT_VALUES); \
618:         } \
619:       } \
620:     } \
621:   }

623: /*
624:  Add column indices into table for counting the nonzeros of merged rows
625:  */
626: #define MatMergeRows_SeqAIJ(mat, nrows, rows, ta) \
627:   { \
628:     PetscInt _j, _row, _nz, *_col, _i; \
629:     for (_i = 0; _i < nrows; _i++) { \
630:       _row = rows[_i]; \
631:       _nz  = mat->i[_row + 1] - mat->i[_row]; \
632:       for (_j = 0; _j < _nz; _j++) { \
633:         _col = _j + mat->j + mat->i[_row]; \
634:         PetscTableAdd(ta, *_col + 1, 1, INSERT_VALUES); \
635:       } \
636:     } \
637:   }

639: #endif