1 
2 #if !defined(__AIJ_H)
3 #define __AIJ_H
4 
5 #include <petsc/private/matimpl.h>
6 #include <petscctable.h>
7 
8 /*
9     Struct header shared by SeqAIJ, SeqBAIJ and SeqSBAIJ matrix formats
10 */
11 #define SEQAIJHEADER(datatype)        \
12   PetscBool roworiented;              /* if true, row-oriented input, default */ \
13   PetscInt  nonew;                    /* 1 don't add new nonzeros, -1 generate error on new */ \
14   PetscInt  nounused;                 /* -1 generate error on unused space */ \
15   PetscBool singlemalloc;             /* if true a, i, and j have been obtained with one big malloc */ \
16   PetscInt  maxnz;                    /* allocated nonzeros */ \
17   PetscInt  *imax;                    /* maximum space allocated for each row */ \
18   PetscInt  *ilen;                    /* actual length of each row */ \
19   PetscInt  *ipre;                    /* space preallocated for each row by user */ \
20   PetscBool free_imax_ilen;  \
21   PetscInt  reallocs;                 /* number of mallocs done during MatSetValues() \
22                                         as more values are set than were prealloced */\
23   PetscInt          rmax;             /* max nonzeros in any row */ \
24   PetscBool         keepnonzeropattern;   /* keeps matrix structure same in calls to MatZeroRows()*/ \
25   PetscBool         ignorezeroentries; \
26   PetscBool         free_ij;          /* free the column indices j and row offsets i when the matrix is destroyed */ \
27   PetscBool         free_a;           /* free the numerical values when matrix is destroy */ \
28   Mat_CompressedRow compressedrow;    /* use compressed row format */                      \
29   PetscInt          nz;               /* nonzeros */                                       \
30   PetscInt          *i;               /* pointer to beginning of each row */               \
31   PetscInt          *j;               /* column values: j + i[k] - 1 is start of row k */  \
32   PetscInt          *diag;            /* pointers to diagonal elements */                  \
33   PetscInt          nonzerorowcnt;    /* how many rows have nonzero entries */             \
34   PetscBool         free_diag;         \
35   datatype          *a;               /* nonzero elements */                               \
36   PetscScalar       *solve_work;      /* work space used in MatSolve */                    \
37   IS                row, col, icol;   /* index sets, used for reorderings */ \
38   PetscBool         pivotinblocks;    /* pivot inside factorization of each diagonal block */ \
39   Mat               parent;           /* set if this matrix was formed with MatDuplicate(...,MAT_SHARE_NONZERO_PATTERN,....); \
40                                          means that this shares some data structures with the parent including diag, ilen, imax, i, j */\
41   Mat_SubSppt       *submatis1         /* used by MatCreateSubMatrices_MPIXAIJ_Local */
42 
43 typedef struct {
44   MatTransposeColoring matcoloring;
45   Mat                  Bt_den;       /* dense matrix of B^T */
46   Mat                  ABt_den;      /* dense matrix of A*B^T */
47   PetscBool            usecoloring;
48 } Mat_MatMatTransMult;
49 
50 typedef struct { /* used by MatTransposeMatMult() */
51   Mat          At;           /* transpose of the first matrix */
52   Mat          mA;           /* maij matrix of A */
53   Vec          bt,ct;        /* vectors to hold locally transposed arrays of B and C */
54   PetscBool    updateAt;     /* flg to avoid recomputing At in MatProductNumeric_AtB_SeqAIJ_SeqAIJ() */
55   /* used by PtAP */
56   void           *data;
57   PetscErrorCode (*destroy)(void*);
58 } Mat_MatTransMatMult;
59 
60 typedef struct {
61   PetscInt    *api,*apj;       /* symbolic structure of A*P */
62   PetscScalar *apa;            /* temporary array for storing one row of A*P */
63 } Mat_AP;
64 
65 typedef struct {
66   MatTransposeColoring matcoloring;
67   Mat                  Rt;    /* sparse or dense matrix of R^T */
68   Mat                  RARt;  /* dense matrix of R*A*R^T */
69   Mat                  ARt;   /* A*R^T used for the case -matrart_color_art */
70   MatScalar            *work; /* work array to store columns of A*R^T used in MatMatMatMultNumeric_SeqAIJ_SeqAIJ_SeqDense() */
71   /* free intermediate products needed for PtAP */
72   void                 *data;
73   PetscErrorCode       (*destroy)(void*);
74 } Mat_RARt;
75 
76 typedef struct {
77   Mat BC;               /* temp matrix for storing B*C */
78 } Mat_MatMatMatMult;
79 
80 /*
81   MATSEQAIJ format - Compressed row storage (also called Yale sparse matrix
82   format) or compressed sparse row (CSR).  The i[] and j[] arrays start at 0. For example,
83   j[i[k]+p] is the pth column in row k.  Note that the diagonal
84   matrix elements are stored with the rest of the nonzeros (not separately).
85 */
86 
87 /* Info about i-nodes (identical nodes) helper class for SeqAIJ */
88 typedef struct {
89   MatScalar        *bdiag,*ibdiag,*ssor_work;        /* diagonal blocks of matrix used for MatSOR_SeqAIJ_Inode() */
90   PetscInt         bdiagsize;                         /* length of bdiag and ibdiag */
91   PetscBool        ibdiagvalid;                       /* do ibdiag[] and bdiag[] contain the most recent values */
92 
93   PetscBool        use;
94   PetscInt         node_count;                     /* number of inodes */
95   PetscInt         *size;                          /* size of each inode */
96   PetscInt         limit;                          /* inode limit */
97   PetscInt         max_limit;                      /* maximum supported inode limit */
98   PetscBool        checked;                        /* if inodes have been checked for */
99   PetscObjectState mat_nonzerostate;               /* non-zero state when inodes were checked for */
100 } Mat_SeqAIJ_Inode;
101 
102 PETSC_INTERN PetscErrorCode MatView_SeqAIJ_Inode(Mat,PetscViewer);
103 PETSC_INTERN PetscErrorCode MatAssemblyEnd_SeqAIJ_Inode(Mat,MatAssemblyType);
104 PETSC_INTERN PetscErrorCode MatDestroy_SeqAIJ_Inode(Mat);
105 PETSC_INTERN PetscErrorCode MatCreate_SeqAIJ_Inode(Mat);
106 PETSC_INTERN PetscErrorCode MatSetOption_SeqAIJ_Inode(Mat,MatOption,PetscBool);
107 PETSC_INTERN PetscErrorCode MatDuplicate_SeqAIJ_Inode(Mat,MatDuplicateOption,Mat*);
108 PETSC_INTERN PetscErrorCode MatDuplicateNoCreate_SeqAIJ(Mat,Mat,MatDuplicateOption,PetscBool);
109 PETSC_INTERN PetscErrorCode MatLUFactorNumeric_SeqAIJ_Inode_inplace(Mat,Mat,const MatFactorInfo*);
110 PETSC_INTERN PetscErrorCode MatLUFactorNumeric_SeqAIJ_Inode(Mat,Mat,const MatFactorInfo*);
111 PETSC_INTERN PetscErrorCode MatSeqAIJGetArray_SeqAIJ(Mat,PetscScalar**);
112 PETSC_INTERN PetscErrorCode MatSeqAIJRestoreArray_SeqAIJ(Mat,PetscScalar**);
113 
114 typedef struct {
115   SEQAIJHEADER(MatScalar);
116   Mat_SeqAIJ_Inode inode;
117   MatScalar        *saved_values;             /* location for stashing nonzero values of matrix */
118 
119   PetscScalar *idiag,*mdiag,*ssor_work;       /* inverse of diagonal entries, diagonal values and workspace for Eisenstat trick */
120   PetscBool   idiagvalid;                     /* current idiag[] and mdiag[] are valid */
121   PetscScalar *ibdiag;                        /* inverses of block diagonals */
122   PetscBool   ibdiagvalid;                    /* inverses of block diagonals are valid. */
123   PetscBool   diagonaldense;                  /* all entries along the diagonal have been set; i.e. no missing diagonal terms */
124   PetscScalar fshift,omega;                   /* last used omega and fshift */
125 } Mat_SeqAIJ;
126 
127 /*
128   Frees the a, i, and j arrays from the XAIJ (AIJ, BAIJ, and SBAIJ) matrix types
129 */
MatSeqXAIJFreeAIJ(Mat AA,MatScalar ** a,PetscInt ** j,PetscInt ** i)130 PETSC_STATIC_INLINE PetscErrorCode MatSeqXAIJFreeAIJ(Mat AA,MatScalar **a,PetscInt **j,PetscInt **i)
131 {
132   PetscErrorCode ierr;
133   Mat_SeqAIJ     *A = (Mat_SeqAIJ*) AA->data;
134   if (A->singlemalloc) {
135     ierr = PetscFree3(*a,*j,*i);CHKERRQ(ierr);
136   } else {
137     if (A->free_a)  {ierr = PetscFree(*a);CHKERRQ(ierr);}
138     if (A->free_ij) {ierr = PetscFree(*j);CHKERRQ(ierr);}
139     if (A->free_ij) {ierr = PetscFree(*i);CHKERRQ(ierr);}
140   }
141   return 0;
142 }
143 /*
144     Allocates larger a, i, and j arrays for the XAIJ (AIJ, BAIJ, and SBAIJ) matrix types
145     This is a macro because it takes the datatype as an argument which can be either a Mat or a MatScalar
146 */
147 #define MatSeqXAIJReallocateAIJ(Amat,AM,BS2,NROW,ROW,COL,RMAX,AA,AI,AJ,RP,AP,AIMAX,NONEW,datatype) \
148   if (NROW >= RMAX) { \
149     Mat_SeqAIJ *Ain = (Mat_SeqAIJ*)Amat->data; \
150     /* there is no extra room in row, therefore enlarge */ \
151     PetscInt CHUNKSIZE = 15,new_nz = AI[AM] + CHUNKSIZE,len,*new_i=NULL,*new_j=NULL; \
152     datatype *new_a; \
153  \
154     if (NONEW == -2) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"New nonzero at (%D,%D) caused a malloc\nUse MatSetOption(A, MAT_NEW_NONZERO_ALLOCATION_ERR, PETSC_FALSE) to turn off this check",ROW,COL); \
155     /* malloc new storage space */ \
156     ierr = PetscMalloc3(BS2*new_nz,&new_a,new_nz,&new_j,AM+1,&new_i);CHKERRQ(ierr); \
157  \
158     /* copy over old data into new slots */ \
159     for (ii=0; ii<ROW+1; ii++) {new_i[ii] = AI[ii];} \
160     for (ii=ROW+1; ii<AM+1; ii++) {new_i[ii] = AI[ii]+CHUNKSIZE;} \
161     ierr = PetscArraycpy(new_j,AJ,AI[ROW]+NROW);CHKERRQ(ierr); \
162     len  = (new_nz - CHUNKSIZE - AI[ROW] - NROW); \
163     ierr = PetscArraycpy(new_j+AI[ROW]+NROW+CHUNKSIZE,AJ+AI[ROW]+NROW,len);CHKERRQ(ierr); \
164     ierr = PetscArraycpy(new_a,AA,BS2*(AI[ROW]+NROW));CHKERRQ(ierr);    \
165     ierr = PetscArrayzero(new_a+BS2*(AI[ROW]+NROW),BS2*CHUNKSIZE);CHKERRQ(ierr); \
166     ierr = PetscArraycpy(new_a+BS2*(AI[ROW]+NROW+CHUNKSIZE),AA+BS2*(AI[ROW]+NROW),BS2*len);CHKERRQ(ierr);  \
167     /* free up old matrix storage */ \
168     ierr              = MatSeqXAIJFreeAIJ(A,&Ain->a,&Ain->j,&Ain->i);CHKERRQ(ierr); \
169     AA                = new_a; \
170     Ain->a            = (MatScalar*) new_a;                   \
171     AI                = Ain->i = new_i; AJ = Ain->j = new_j;  \
172     Ain->singlemalloc = PETSC_TRUE; \
173  \
174     RP          = AJ + AI[ROW]; AP = AA + BS2*AI[ROW]; \
175     RMAX        = AIMAX[ROW] = AIMAX[ROW] + CHUNKSIZE; \
176     Ain->maxnz += BS2*CHUNKSIZE; \
177     Ain->reallocs++; \
178   } \
179 
180 #define MatSeqXAIJReallocateAIJ_structure_only(Amat,AM,BS2,NROW,ROW,COL,RMAX,AI,AJ,RP,AIMAX,NONEW,datatype) \
181   if (NROW >= RMAX) { \
182     Mat_SeqAIJ *Ain = (Mat_SeqAIJ*)Amat->data; \
183     /* there is no extra room in row, therefore enlarge */ \
184     PetscInt CHUNKSIZE = 15,new_nz = AI[AM] + CHUNKSIZE,len,*new_i=NULL,*new_j=NULL; \
185  \
186     if (NONEW == -2) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"New nonzero at (%D,%D) caused a malloc\nUse MatSetOption(A, MAT_NEW_NONZERO_ALLOCATION_ERR, PETSC_FALSE) to turn off this check",ROW,COL); \
187     /* malloc new storage space */ \
188     ierr = PetscMalloc1(new_nz,&new_j);CHKERRQ(ierr); \
189     ierr = PetscMalloc1(AM+1,&new_i);CHKERRQ(ierr);\
190  \
191     /* copy over old data into new slots */ \
192     for (ii=0; ii<ROW+1; ii++) {new_i[ii] = AI[ii];} \
193     for (ii=ROW+1; ii<AM+1; ii++) {new_i[ii] = AI[ii]+CHUNKSIZE;} \
194     ierr = PetscArraycpy(new_j,AJ,AI[ROW]+NROW);CHKERRQ(ierr); \
195     len  = (new_nz - CHUNKSIZE - AI[ROW] - NROW); \
196     ierr = PetscArraycpy(new_j+AI[ROW]+NROW+CHUNKSIZE,AJ+AI[ROW]+NROW,len);CHKERRQ(ierr); \
197  \
198     /* free up old matrix storage */ \
199     ierr              = MatSeqXAIJFreeAIJ(A,&Ain->a,&Ain->j,&Ain->i);CHKERRQ(ierr); \
200     Ain->a            = NULL;                   \
201     AI                = Ain->i = new_i; AJ = Ain->j = new_j;  \
202     Ain->singlemalloc = PETSC_FALSE; \
203     Ain->free_a       = PETSC_FALSE;             \
204  \
205     RP          = AJ + AI[ROW];    \
206     RMAX        = AIMAX[ROW] = AIMAX[ROW] + CHUNKSIZE; \
207     Ain->maxnz += BS2*CHUNKSIZE; \
208     Ain->reallocs++; \
209   } \
210 
211 PETSC_INTERN PetscErrorCode MatSeqAIJSetPreallocation_SeqAIJ(Mat,PetscInt,const PetscInt*);
212 PETSC_INTERN PetscErrorCode MatILUFactorSymbolic_SeqAIJ_inplace(Mat,Mat,IS,IS,const MatFactorInfo*);
213 PETSC_INTERN PetscErrorCode MatILUFactorSymbolic_SeqAIJ(Mat,Mat,IS,IS,const MatFactorInfo*);
214 PETSC_INTERN PetscErrorCode MatILUFactorSymbolic_SeqAIJ_ilu0(Mat,Mat,IS,IS,const MatFactorInfo*);
215 
216 PETSC_INTERN PetscErrorCode MatICCFactorSymbolic_SeqAIJ_inplace(Mat,Mat,IS,const MatFactorInfo*);
217 PETSC_INTERN PetscErrorCode MatICCFactorSymbolic_SeqAIJ(Mat,Mat,IS,const MatFactorInfo*);
218 PETSC_INTERN PetscErrorCode MatCholeskyFactorSymbolic_SeqAIJ_inplace(Mat,Mat,IS,const MatFactorInfo*);
219 PETSC_INTERN PetscErrorCode MatCholeskyFactorSymbolic_SeqAIJ(Mat,Mat,IS,const MatFactorInfo*);
220 PETSC_INTERN PetscErrorCode MatCholeskyFactorNumeric_SeqAIJ_inplace(Mat,Mat,const MatFactorInfo*);
221 PETSC_INTERN PetscErrorCode MatCholeskyFactorNumeric_SeqAIJ(Mat,Mat,const MatFactorInfo*);
222 PETSC_INTERN PetscErrorCode MatDuplicate_SeqAIJ(Mat,MatDuplicateOption,Mat*);
223 PETSC_INTERN PetscErrorCode MatCopy_SeqAIJ(Mat,Mat,MatStructure);
224 PETSC_INTERN PetscErrorCode MatMissingDiagonal_SeqAIJ(Mat,PetscBool*,PetscInt*);
225 PETSC_INTERN PetscErrorCode MatMarkDiagonal_SeqAIJ(Mat);
226 PETSC_INTERN PetscErrorCode MatFindZeroDiagonals_SeqAIJ_Private(Mat,PetscInt*,PetscInt**);
227 
228 PETSC_INTERN PetscErrorCode MatMult_SeqAIJ(Mat,Vec,Vec);
229 PETSC_INTERN PetscErrorCode MatMult_SeqAIJ_Inode(Mat,Vec,Vec);
230 PETSC_INTERN PetscErrorCode MatMultAdd_SeqAIJ(Mat,Vec,Vec,Vec);
231 PETSC_INTERN PetscErrorCode MatMultAdd_SeqAIJ_Inode(Mat,Vec,Vec,Vec);
232 PETSC_INTERN PetscErrorCode MatMultTranspose_SeqAIJ(Mat,Vec,Vec);
233 PETSC_INTERN PetscErrorCode MatMultTransposeAdd_SeqAIJ(Mat,Vec,Vec,Vec);
234 PETSC_INTERN PetscErrorCode MatSOR_SeqAIJ(Mat,Vec,PetscReal,MatSORType,PetscReal,PetscInt,PetscInt,Vec);
235 PETSC_INTERN PetscErrorCode MatSOR_SeqAIJ_Inode(Mat,Vec,PetscReal,MatSORType,PetscReal,PetscInt,PetscInt,Vec);
236 
237 PETSC_INTERN PetscErrorCode MatSetOption_SeqAIJ(Mat,MatOption,PetscBool);
238 
239 PETSC_INTERN PetscErrorCode MatGetSymbolicTranspose_SeqAIJ(Mat,PetscInt *[],PetscInt *[]);
240 PETSC_INTERN PetscErrorCode MatGetSymbolicTransposeReduced_SeqAIJ(Mat,PetscInt,PetscInt,PetscInt *[],PetscInt *[]);
241 PETSC_INTERN PetscErrorCode MatRestoreSymbolicTranspose_SeqAIJ(Mat,PetscInt *[],PetscInt *[]);
242 PETSC_INTERN PetscErrorCode MatTransposeSymbolic_SeqAIJ(Mat,Mat*);
243 PETSC_INTERN PetscErrorCode MatTranspose_SeqAIJ(Mat,MatReuse,Mat*);
244 PETSC_INTERN PetscErrorCode MatToSymmetricIJ_SeqAIJ(PetscInt,PetscInt*,PetscInt*,PetscBool,PetscInt,PetscInt,PetscInt**,PetscInt**);
245 PETSC_INTERN PetscErrorCode MatLUFactorSymbolic_SeqAIJ_inplace(Mat,Mat,IS,IS,const MatFactorInfo*);
246 PETSC_INTERN PetscErrorCode MatLUFactorSymbolic_SeqAIJ(Mat,Mat,IS,IS,const MatFactorInfo*);
247 PETSC_INTERN PetscErrorCode MatLUFactorNumeric_SeqAIJ_inplace(Mat,Mat,const MatFactorInfo*);
248 PETSC_INTERN PetscErrorCode MatLUFactorNumeric_SeqAIJ(Mat,Mat,const MatFactorInfo*);
249 PETSC_INTERN PetscErrorCode MatLUFactorNumeric_SeqAIJ_InplaceWithPerm(Mat,Mat,const MatFactorInfo*);
250 PETSC_INTERN PetscErrorCode MatLUFactor_SeqAIJ(Mat,IS,IS,const MatFactorInfo*);
251 PETSC_INTERN PetscErrorCode MatSolve_SeqAIJ_inplace(Mat,Vec,Vec);
252 PETSC_INTERN PetscErrorCode MatSolve_SeqAIJ(Mat,Vec,Vec);
253 PETSC_INTERN PetscErrorCode MatSolve_SeqAIJ_Inode_inplace(Mat,Vec,Vec);
254 PETSC_INTERN PetscErrorCode MatSolve_SeqAIJ_Inode(Mat,Vec,Vec);
255 PETSC_INTERN PetscErrorCode MatSolve_SeqAIJ_NaturalOrdering_inplace(Mat,Vec,Vec);
256 PETSC_INTERN PetscErrorCode MatSolve_SeqAIJ_NaturalOrdering(Mat,Vec,Vec);
257 PETSC_INTERN PetscErrorCode MatSolve_SeqAIJ_InplaceWithPerm(Mat,Vec,Vec);
258 PETSC_INTERN PetscErrorCode MatSolveAdd_SeqAIJ_inplace(Mat,Vec,Vec,Vec);
259 PETSC_INTERN PetscErrorCode MatSolveAdd_SeqAIJ(Mat,Vec,Vec,Vec);
260 PETSC_INTERN PetscErrorCode MatSolveTranspose_SeqAIJ_inplace(Mat,Vec,Vec);
261 PETSC_INTERN PetscErrorCode MatSolveTranspose_SeqAIJ(Mat,Vec,Vec);
262 PETSC_INTERN PetscErrorCode MatSolveTransposeAdd_SeqAIJ_inplace(Mat,Vec,Vec,Vec);
263 PETSC_INTERN PetscErrorCode MatSolveTransposeAdd_SeqAIJ(Mat,Vec,Vec,Vec);
264 PETSC_INTERN PetscErrorCode MatMatSolve_SeqAIJ_inplace(Mat,Mat,Mat);
265 PETSC_INTERN PetscErrorCode MatMatSolve_SeqAIJ(Mat,Mat,Mat);
266 PETSC_INTERN PetscErrorCode MatEqual_SeqAIJ(Mat,Mat,PetscBool*);
267 PETSC_INTERN PetscErrorCode MatFDColoringCreate_SeqXAIJ(Mat,ISColoring,MatFDColoring);
268 PETSC_INTERN PetscErrorCode MatFDColoringSetUp_SeqXAIJ(Mat,ISColoring,MatFDColoring);
269 PETSC_INTERN PetscErrorCode MatFDColoringSetUpBlocked_AIJ_Private(Mat,MatFDColoring,PetscInt);
270 PETSC_INTERN PetscErrorCode MatLoad_AIJ_HDF5(Mat,PetscViewer);
271 PETSC_INTERN PetscErrorCode MatLoad_SeqAIJ_Binary(Mat,PetscViewer);
272 PETSC_INTERN PetscErrorCode MatLoad_SeqAIJ(Mat,PetscViewer);
273 PETSC_INTERN PetscErrorCode RegisterApplyPtAPRoutines_Private(Mat);
274 
275 #if defined(PETSC_HAVE_HYPRE)
276 PETSC_INTERN PetscErrorCode MatProductSetFromOptions_Transpose_AIJ_AIJ(Mat);
277 #endif
278 PETSC_INTERN PetscErrorCode MatProductSetFromOptions_SeqAIJ(Mat);
279 
280 PETSC_INTERN PetscErrorCode MatProductSymbolic_SeqAIJ_SeqAIJ(Mat);
281 PETSC_INTERN PetscErrorCode MatProductSymbolic_PtAP_SeqAIJ_SeqAIJ(Mat);
282 PETSC_INTERN PetscErrorCode MatProductSymbolic_RARt_SeqAIJ_SeqAIJ(Mat);
283 
284 PETSC_INTERN PetscErrorCode MatMatMultSymbolic_SeqAIJ_SeqAIJ(Mat,Mat,PetscReal,Mat);
285 PETSC_INTERN PetscErrorCode MatMatMultSymbolic_SeqAIJ_SeqAIJ_Sorted(Mat,Mat,PetscReal,Mat);
286 PETSC_INTERN PetscErrorCode MatMatMultSymbolic_SeqDense_SeqAIJ(Mat,Mat,PetscReal,Mat);
287 PETSC_INTERN PetscErrorCode MatMatMultSymbolic_SeqAIJ_SeqAIJ_Scalable(Mat,Mat,PetscReal,Mat);
288 PETSC_INTERN PetscErrorCode MatMatMultSymbolic_SeqAIJ_SeqAIJ_Scalable_fast(Mat,Mat,PetscReal,Mat);
289 PETSC_INTERN PetscErrorCode MatMatMultSymbolic_SeqAIJ_SeqAIJ_Heap(Mat,Mat,PetscReal,Mat);
290 PETSC_INTERN PetscErrorCode MatMatMultSymbolic_SeqAIJ_SeqAIJ_BTHeap(Mat,Mat,PetscReal,Mat);
291 PETSC_INTERN PetscErrorCode MatMatMultSymbolic_SeqAIJ_SeqAIJ_RowMerge(Mat,Mat,PetscReal,Mat);
292 PETSC_INTERN PetscErrorCode MatMatMultSymbolic_SeqAIJ_SeqAIJ_LLCondensed(Mat,Mat,PetscReal,Mat);
293 #if defined(PETSC_HAVE_HYPRE)
294 PETSC_INTERN PetscErrorCode MatMatMultSymbolic_AIJ_AIJ_wHYPRE(Mat,Mat,PetscReal,Mat);
295 #endif
296 
297 PETSC_INTERN PetscErrorCode MatMatMultNumeric_SeqAIJ_SeqAIJ(Mat,Mat,Mat);
298 PETSC_INTERN PetscErrorCode MatMatMultNumeric_SeqAIJ_SeqAIJ_Sorted(Mat,Mat,Mat);
299 
300 PETSC_INTERN PetscErrorCode MatMatMultNumeric_SeqDense_SeqAIJ(Mat,Mat,Mat);
301 PETSC_INTERN PetscErrorCode MatMatMultNumeric_SeqAIJ_SeqAIJ_Scalable(Mat,Mat,Mat);
302 
303 PETSC_INTERN PetscErrorCode MatPtAPSymbolic_SeqAIJ_SeqAIJ_SparseAxpy(Mat,Mat,PetscReal,Mat);
304 PETSC_INTERN PetscErrorCode MatPtAPNumeric_SeqAIJ_SeqAIJ(Mat,Mat,Mat);
305 PETSC_INTERN PetscErrorCode MatPtAPNumeric_SeqAIJ_SeqAIJ_SparseAxpy(Mat,Mat,Mat);
306 
307 PETSC_INTERN PetscErrorCode MatRARtSymbolic_SeqAIJ_SeqAIJ(Mat,Mat,PetscReal,Mat);
308 PETSC_INTERN PetscErrorCode MatRARtSymbolic_SeqAIJ_SeqAIJ_matmattransposemult(Mat,Mat,PetscReal,Mat);
309 PETSC_INTERN PetscErrorCode MatRARtSymbolic_SeqAIJ_SeqAIJ_colorrart(Mat,Mat,PetscReal,Mat);
310 PETSC_INTERN PetscErrorCode MatRARtNumeric_SeqAIJ_SeqAIJ(Mat,Mat,Mat);
311 PETSC_INTERN PetscErrorCode MatRARtNumeric_SeqAIJ_SeqAIJ_matmattransposemult(Mat,Mat,Mat);
312 PETSC_INTERN PetscErrorCode MatRARtNumeric_SeqAIJ_SeqAIJ_colorrart(Mat,Mat,Mat);
313 
314 PETSC_INTERN PetscErrorCode MatTransposeMatMultSymbolic_SeqAIJ_SeqAIJ(Mat,Mat,PetscReal,Mat);
315 PETSC_INTERN PetscErrorCode MatTransposeMatMultNumeric_SeqAIJ_SeqAIJ(Mat,Mat,Mat);
316 PETSC_INTERN PetscErrorCode MatDestroy_SeqAIJ_MatTransMatMult(void*);
317 
318 PETSC_INTERN PetscErrorCode MatMatTransposeMultSymbolic_SeqAIJ_SeqAIJ(Mat,Mat,PetscReal,Mat);
319 PETSC_INTERN PetscErrorCode MatMatTransposeMultNumeric_SeqAIJ_SeqAIJ(Mat,Mat,Mat);
320 PETSC_INTERN PetscErrorCode MatTransposeColoringCreate_SeqAIJ(Mat,ISColoring,MatTransposeColoring);
321 PETSC_INTERN PetscErrorCode MatTransColoringApplySpToDen_SeqAIJ(MatTransposeColoring,Mat,Mat);
322 PETSC_INTERN PetscErrorCode MatTransColoringApplyDenToSp_SeqAIJ(MatTransposeColoring,Mat,Mat);
323 
324 PETSC_INTERN PetscErrorCode MatMatMatMultSymbolic_SeqAIJ_SeqAIJ_SeqAIJ(Mat,Mat,Mat,PetscReal,Mat);
325 PETSC_INTERN PetscErrorCode MatMatMatMultNumeric_SeqAIJ_SeqAIJ_SeqAIJ(Mat,Mat,Mat,Mat);
326 
327 PETSC_INTERN PetscErrorCode MatSetRandomSkipColumnRange_SeqAIJ_Private(Mat,PetscInt,PetscInt,PetscRandom);
328 PETSC_INTERN PetscErrorCode MatSetValues_SeqAIJ(Mat,PetscInt,const PetscInt[],PetscInt,const PetscInt[],const PetscScalar[],InsertMode);
329 PETSC_INTERN PetscErrorCode MatGetRow_SeqAIJ(Mat,PetscInt,PetscInt*,PetscInt**,PetscScalar**);
330 PETSC_INTERN PetscErrorCode MatRestoreRow_SeqAIJ(Mat,PetscInt,PetscInt*,PetscInt**,PetscScalar**);
331 PETSC_INTERN PetscErrorCode MatScale_SeqAIJ(Mat,PetscScalar);
332 PETSC_INTERN PetscErrorCode MatDiagonalScale_SeqAIJ(Mat,Vec,Vec);
333 PETSC_INTERN PetscErrorCode MatDiagonalSet_SeqAIJ(Mat,Vec,InsertMode);
334 PETSC_INTERN PetscErrorCode MatAXPY_SeqAIJ(Mat,PetscScalar,Mat,MatStructure);
335 PETSC_INTERN PetscErrorCode MatGetRowIJ_SeqAIJ(Mat,PetscInt,PetscBool,PetscBool,PetscInt*,const PetscInt *[],const PetscInt *[],PetscBool*);
336 PETSC_INTERN PetscErrorCode MatRestoreRowIJ_SeqAIJ(Mat,PetscInt,PetscBool,PetscBool,PetscInt*,const PetscInt *[],const PetscInt *[],PetscBool*);
337 PETSC_INTERN PetscErrorCode MatGetColumnIJ_SeqAIJ(Mat,PetscInt,PetscBool,PetscBool,PetscInt*,const PetscInt *[],const PetscInt *[],PetscBool*);
338 PETSC_INTERN PetscErrorCode MatRestoreColumnIJ_SeqAIJ(Mat,PetscInt,PetscBool,PetscBool,PetscInt*,const PetscInt *[],const PetscInt *[],PetscBool*);
339 PETSC_INTERN PetscErrorCode MatGetColumnIJ_SeqAIJ_Color(Mat,PetscInt,PetscBool,PetscBool,PetscInt*,const PetscInt *[],const PetscInt *[],PetscInt *[],PetscBool*);
340 PETSC_INTERN PetscErrorCode MatRestoreColumnIJ_SeqAIJ_Color(Mat,PetscInt,PetscBool,PetscBool,PetscInt*,const PetscInt *[],const PetscInt *[],PetscInt *[],PetscBool*);
341 PETSC_INTERN PetscErrorCode MatDestroy_SeqAIJ(Mat);
342 PETSC_INTERN PetscErrorCode MatSetUp_SeqAIJ(Mat);
343 PETSC_INTERN PetscErrorCode MatView_SeqAIJ(Mat,PetscViewer);
344 
345 PETSC_INTERN PetscErrorCode MatSeqAIJInvalidateDiagonal(Mat);
346 PETSC_INTERN PetscErrorCode MatSeqAIJInvalidateDiagonal_Inode(Mat);
347 PETSC_INTERN PetscErrorCode MatSeqAIJCheckInode(Mat);
348 PETSC_INTERN PetscErrorCode MatSeqAIJCheckInode_FactorLU(Mat);
349 
350 PETSC_INTERN PetscErrorCode MatAXPYGetPreallocation_SeqAIJ(Mat,Mat,PetscInt*);
351 
352 #if defined(PETSC_HAVE_MATLAB_ENGINE)
353 PETSC_EXTERN PetscErrorCode MatlabEnginePut_SeqAIJ(PetscObject,void*);
354 PETSC_EXTERN PetscErrorCode MatlabEngineGet_SeqAIJ(PetscObject,void*);
355 #endif
356 PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqSBAIJ(Mat,MatType,MatReuse,Mat*);
357 PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqBAIJ(Mat,MatType,MatReuse,Mat*);
358 PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqDense(Mat,MatType,MatReuse,Mat*);
359 PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqAIJCRL(Mat,MatType,MatReuse,Mat*);
360 PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_Elemental(Mat,MatType,MatReuse,Mat*);
361 #if defined(PETSC_HAVE_SCALAPACK)
362 PETSC_INTERN PetscErrorCode MatConvert_AIJ_ScaLAPACK(Mat,MatType,MatReuse,Mat*);
363 #endif
364 PETSC_INTERN PetscErrorCode MatConvert_AIJ_HYPRE(Mat,MatType,MatReuse,Mat*);
365 PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqAIJPERM(Mat,MatType,MatReuse,Mat*);
366 PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqAIJSELL(Mat,MatType,MatReuse,Mat*);
367 PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqAIJMKL(Mat,MatType,MatReuse,Mat*);
368 PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqAIJViennaCL(Mat,MatType,MatReuse,Mat*);
369 PETSC_INTERN PetscErrorCode MatReorderForNonzeroDiagonal_SeqAIJ(Mat,PetscReal,IS,IS);
370 PETSC_INTERN PetscErrorCode MatRARt_SeqAIJ_SeqAIJ(Mat,Mat,MatReuse,PetscReal,Mat*);
371 PETSC_EXTERN PetscErrorCode MatCreate_SeqAIJ(Mat);
372 PETSC_INTERN PetscErrorCode MatAssemblyEnd_SeqAIJ(Mat,MatAssemblyType);
373 PETSC_EXTERN PetscErrorCode MatZeroEntries_SeqAIJ(Mat);
374 
375 PETSC_INTERN PetscErrorCode MatAXPYGetPreallocation_SeqX_private(PetscInt,const PetscInt*,const PetscInt*,const PetscInt*,const PetscInt*,PetscInt*);
376 PETSC_INTERN PetscErrorCode MatCreateMPIMatConcatenateSeqMat_SeqAIJ(MPI_Comm,Mat,PetscInt,MatReuse,Mat*);
377 PETSC_INTERN PetscErrorCode MatCreateMPIMatConcatenateSeqMat_MPIAIJ(MPI_Comm,Mat,PetscInt,MatReuse,Mat*);
378 
379 PETSC_INTERN PetscErrorCode MatSetSeqMat_SeqAIJ(Mat,IS,IS,MatStructure,Mat);
380 PETSC_INTERN PetscErrorCode MatDestroySubMatrix_Private(Mat_SubSppt*);
381 PETSC_INTERN PetscErrorCode MatDestroySubMatrix_SeqAIJ(Mat);
382 PETSC_INTERN PetscErrorCode MatDestroySubMatrix_Dummy(Mat);
383 PETSC_INTERN PetscErrorCode MatDestroySubMatrices_Dummy(PetscInt, Mat*[]);
384 PETSC_INTERN PetscErrorCode MatCreateSubMatrix_SeqAIJ(Mat,IS,IS,PetscInt,MatReuse,Mat*);
385 
386 PETSC_INTERN PetscErrorCode MatSeqAIJCompactOutExtraColumns_SeqAIJ(Mat,ISLocalToGlobalMapping*);
387 PETSC_INTERN PetscErrorCode MatSetSeqAIJWithArrays_private(MPI_Comm,PetscInt,PetscInt,PetscInt[],PetscInt[],PetscScalar[],MatType,Mat);
388 
389 /*
390     PetscSparseDenseMinusDot - The inner kernel of triangular solves and Gauss-Siedel smoothing. \sum_i xv[i] * r[xi[i]] for CSR storage
391 
392   Input Parameters:
393 +  nnz - the number of entries
394 .  r - the array of vector values
395 .  xv - the matrix values for the row
396 -  xi - the column indices of the nonzeros in the row
397 
398   Output Parameter:
399 .  sum - negative the sum of results
400 
401   PETSc compile flags:
402 +   PETSC_KERNEL_USE_UNROLL_4
403 -   PETSC_KERNEL_USE_UNROLL_2
404 
405   Developer Notes:
406     The macro changes sum but not other parameters
407 
408 .seealso: PetscSparseDensePlusDot()
409 
410 */
411 #if defined(PETSC_KERNEL_USE_UNROLL_4)
412 #define PetscSparseDenseMinusDot(sum,r,xv,xi,nnz) { \
413     if (nnz > 0) { \
414       PetscInt nnz2=nnz,rem=nnz&0x3; \
415       switch (rem) { \
416       case 3: sum -= *xv++ *r[*xi++]; \
417       case 2: sum -= *xv++ *r[*xi++]; \
418       case 1: sum -= *xv++ *r[*xi++]; \
419         nnz2      -= rem;} \
420       while (nnz2 > 0) { \
421         sum -=  xv[0] * r[xi[0]] + xv[1] * r[xi[1]] + \
422                 xv[2] * r[xi[2]] + xv[3] * r[xi[3]]; \
423         xv += 4; xi += 4; nnz2 -= 4; \
424       } \
425       xv -= nnz; xi -= nnz; \
426     } \
427   }
428 
429 #elif defined(PETSC_KERNEL_USE_UNROLL_2)
430 #define PetscSparseDenseMinusDot(sum,r,xv,xi,nnz) { \
431     PetscInt __i,__i1,__i2; \
432     for (__i=0; __i<nnz-1; __i+=2) {__i1 = xi[__i]; __i2=xi[__i+1]; \
433                                     sum -= (xv[__i]*r[__i1] + xv[__i+1]*r[__i2]);} \
434     if (nnz & 0x1) sum -= xv[__i] * r[xi[__i]];}
435 
436 #else
437 #define PetscSparseDenseMinusDot(sum,r,xv,xi,nnz) { \
438     PetscInt __i; \
439     for (__i=0; __i<nnz; __i++) sum -= xv[__i] * r[xi[__i]];}
440 #endif
441 
442 
443 
444 /*
445     PetscSparseDensePlusDot - The inner kernel of matrix-vector product \sum_i xv[i] * r[xi[i]] for CSR storage
446 
447   Input Parameters:
448 +  nnz - the number of entries
449 .  r - the array of vector values
450 .  xv - the matrix values for the row
451 -  xi - the column indices of the nonzeros in the row
452 
453   Output Parameter:
454 .  sum - the sum of results
455 
456   PETSc compile flags:
457 +   PETSC_KERNEL_USE_UNROLL_4
458 -   PETSC_KERNEL_USE_UNROLL_2
459 
460   Developer Notes:
461     The macro changes sum but not other parameters
462 
463 .seealso: PetscSparseDenseMinusDot()
464 
465 */
466 #if defined(PETSC_KERNEL_USE_UNROLL_4)
467 #define PetscSparseDensePlusDot(sum,r,xv,xi,nnz) { \
468     if (nnz > 0) { \
469       PetscInt nnz2=nnz,rem=nnz&0x3; \
470       switch (rem) { \
471       case 3: sum += *xv++ *r[*xi++]; \
472       case 2: sum += *xv++ *r[*xi++]; \
473       case 1: sum += *xv++ *r[*xi++]; \
474         nnz2      -= rem;} \
475       while (nnz2 > 0) { \
476         sum +=  xv[0] * r[xi[0]] + xv[1] * r[xi[1]] + \
477                 xv[2] * r[xi[2]] + xv[3] * r[xi[3]]; \
478         xv += 4; xi += 4; nnz2 -= 4; \
479       } \
480       xv -= nnz; xi -= nnz; \
481     } \
482   }
483 
484 #elif defined(PETSC_KERNEL_USE_UNROLL_2)
485 #define PetscSparseDensePlusDot(sum,r,xv,xi,nnz) { \
486     PetscInt __i,__i1,__i2; \
487     for (__i=0; __i<nnz-1; __i+=2) {__i1 = xi[__i]; __i2=xi[__i+1]; \
488                                     sum += (xv[__i]*r[__i1] + xv[__i+1]*r[__i2]);} \
489     if (nnz & 0x1) sum += xv[__i] * r[xi[__i]];}
490 
491 #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)
492 #define PetscSparseDensePlusDot(sum,r,xv,xi,nnz) PetscSparseDensePlusDot_AVX512_Private(&(sum),(r),(xv),(xi),(nnz))
493 
494 #else
495 #define PetscSparseDensePlusDot(sum,r,xv,xi,nnz) { \
496     PetscInt __i; \
497     for (__i=0; __i<nnz; __i++) sum += xv[__i] * r[xi[__i]];}
498 #endif
499 
500 #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)
501   #include <immintrin.h>
502   #if !defined(_MM_SCALE_8)
503   #define _MM_SCALE_8    8
504   #endif
505 
PetscSparseDensePlusDot_AVX512_Private(PetscScalar * sum,const PetscScalar * x,const MatScalar * aa,const PetscInt * aj,PetscInt n)506 PETSC_STATIC_INLINE void PetscSparseDensePlusDot_AVX512_Private(PetscScalar *sum,const PetscScalar *x,const MatScalar *aa,const PetscInt *aj,PetscInt n)
507 {
508   __m512d  vec_x,vec_y,vec_vals;
509   __m256i  vec_idx;
510   __mmask8 mask;
511   PetscInt j;
512 
513   vec_y = _mm512_setzero_pd();
514   for (j=0; j<(n>>3); j++) {
515     vec_idx  = _mm256_loadu_si256((__m256i const*)aj);
516     vec_vals = _mm512_loadu_pd(aa);
517     vec_x    = _mm512_i32gather_pd(vec_idx,x,_MM_SCALE_8);
518     vec_y    = _mm512_fmadd_pd(vec_x,vec_vals,vec_y);
519     aj += 8; aa += 8;
520   }
521   /* masked load does not work on KNL, it requires avx512vl */
522   if ((n&0x07)>2) {
523     mask     = (__mmask8)(0xff >> (8-(n&0x07)));
524     vec_idx  = _mm256_loadu_si256((__m256i const*)aj);
525     vec_vals = _mm512_loadu_pd(aa);
526     vec_x    = _mm512_mask_i32gather_pd(vec_x,mask,vec_idx,x,_MM_SCALE_8);
527     vec_y    = _mm512_mask3_fmadd_pd(vec_x,vec_vals,vec_y,mask);
528   } else if ((n&0x07)==2) {
529     *sum += aa[0]*x[aj[0]];
530     *sum += aa[1]*x[aj[1]];
531   } else if ((n&0x07)==1) {
532     *sum += aa[0]*x[aj[0]];
533   }
534   if (n>2) *sum += _mm512_reduce_add_pd(vec_y);
535 /*
536   for (j=0;j<(n&0x07);j++) *sum += aa[j]*x[aj[j]];
537 */
538 }
539 #endif
540 
541 /*
542     PetscSparseDenseMaxDot - The inner kernel of a modified matrix-vector product \max_i xv[i] * r[xi[i]] for CSR storage
543 
544   Input Parameters:
545 +  nnz - the number of entries
546 .  r - the array of vector values
547 .  xv - the matrix values for the row
548 -  xi - the column indices of the nonzeros in the row
549 
550   Output Parameter:
551 .  max - the max of results
552 
553 .seealso: PetscSparseDensePlusDot(), PetscSparseDenseMinusDot()
554 
555 */
556 #define PetscSparseDenseMaxDot(max,r,xv,xi,nnz) { \
557     PetscInt __i; \
558     for (__i=0; __i<nnz; __i++) max = PetscMax(PetscRealPart(max), PetscRealPart(xv[__i] * r[xi[__i]]));}
559 
560 /*
561  Add column indices into table for counting the max nonzeros of merged rows
562  */
563 #define MatRowMergeMax_SeqAIJ(mat,nrows,ta) {       \
564     PetscInt _j,_row,_nz,*_col;                     \
565     if (mat) { \
566       for (_row=0; _row<nrows; _row++) {   \
567         _nz = mat->i[_row+1] - mat->i[_row];    \
568         for (_j=0; _j<_nz; _j++) {               \
569           _col = _j + mat->j + mat->i[_row];       \
570           PetscTableAdd(ta,*_col+1,1,INSERT_VALUES);                    \
571         }                                                               \
572       }                                                                 \
573     }    \
574 }
575 
576 /*
577  Add column indices into table for counting the nonzeros of merged rows
578  */
579 #define MatMergeRows_SeqAIJ(mat,nrows,rows,ta) {    \
580   PetscInt _j,_row,_nz,*_col,_i;                      \
581     for (_i=0; _i<nrows; _i++) {\
582       _row = rows[_i]; \
583       _nz = mat->i[_row+1] - mat->i[_row]; \
584       for (_j=0; _j<_nz; _j++) {                \
585         _col = _j + mat->j + mat->i[_row];       \
586         PetscTableAdd(ta,*_col+1,1,INSERT_VALUES); \
587       }                                                                 \
588     }                                                                   \
589 }
590 
591 #endif
592