Actual source code: aij.h
1: #pragma once
3: #include <petsc/private/matimpl.h>
4: #include <petsc/private/hashmapi.h>
5: #include <petsc/private/hashmapijv.h>
7: /*
8: Used by MatCreateSubMatrices_MPIXAIJ_Local()
9: */
10: typedef struct { /* used by MatCreateSubMatrices_MPIAIJ_SingleIS_Local() and MatCreateSubMatrices_MPIAIJ_Local */
11: PetscInt id; /* index of submats, only submats[0] is responsible for deleting some arrays below */
12: PetscInt nrqs, nrqr;
13: PetscInt **rbuf1, **rbuf2, **rbuf3, **sbuf1, **sbuf2;
14: PetscInt **ptr;
15: PetscInt *tmp;
16: PetscInt *ctr;
17: PetscInt *pa; /* proc array */
18: PetscInt *req_size, *req_source1, *req_source2;
19: PetscBool allcolumns, allrows;
20: PetscBool singleis;
21: PetscInt *row2proc; /* row to proc map */
22: PetscInt nstages;
23: #if defined(PETSC_USE_CTABLE)
24: PetscHMapI cmap, rmap;
25: PetscInt *cmap_loc, *rmap_loc;
26: #else
27: PetscInt *cmap, *rmap;
28: #endif
29: PetscErrorCode (*destroy)(Mat);
30: } Mat_SubSppt;
32: /* Operations provided by MATSEQAIJ and its subclasses */
33: typedef struct {
34: PetscErrorCode (*getarray)(Mat, PetscScalar **);
35: PetscErrorCode (*restorearray)(Mat, PetscScalar **);
36: PetscErrorCode (*getarrayread)(Mat, const PetscScalar **);
37: PetscErrorCode (*restorearrayread)(Mat, const PetscScalar **);
38: PetscErrorCode (*getarraywrite)(Mat, PetscScalar **);
39: PetscErrorCode (*restorearraywrite)(Mat, PetscScalar **);
40: PetscErrorCode (*getcsrandmemtype)(Mat, const PetscInt **, const PetscInt **, PetscScalar **, PetscMemType *);
41: } Mat_SeqAIJOps;
43: /*
44: Struct header shared by SeqAIJ, SeqBAIJ and SeqSBAIJ matrix formats
45: */
46: #define SEQAIJHEADER(datatype) \
47: PetscBool roworiented; /* if true, row-oriented input, default */ \
48: PetscInt nonew; /* 1 don't add new nonzeros, -1 generate error on new */ \
49: PetscInt nounused; /* -1 generate error on unused space */ \
50: PetscBool singlemalloc; /* if true a, i, and j have been obtained with one big malloc */ \
51: PetscInt maxnz; /* allocated nonzeros */ \
52: PetscInt *imax; /* maximum space allocated for each row */ \
53: PetscInt *ilen; /* actual length of each row */ \
54: PetscInt *ipre; /* space preallocated for each row by user */ \
55: PetscBool free_imax_ilen; \
56: PetscInt reallocs; /* number of mallocs done during MatSetValues() \
57: as more values are set than were prealloced */ \
58: PetscInt rmax; /* max nonzeros in any row */ \
59: PetscBool keepnonzeropattern; /* keeps matrix structure same in calls to MatZeroRows()*/ \
60: PetscBool ignorezeroentries; \
61: PetscBool free_ij; /* free the column indices j and row offsets i when the matrix is destroyed */ \
62: PetscBool free_a; /* free the numerical values when matrix is destroy */ \
63: Mat_CompressedRow compressedrow; /* use compressed row format */ \
64: PetscInt nz; /* nonzeros */ \
65: PetscInt *i; /* pointer to beginning of each row */ \
66: PetscInt *j; /* column values: j + i[k] - 1 is start of row k */ \
67: PetscInt *diag; /* pointers to diagonal elements */ \
68: PetscInt nonzerorowcnt; /* how many rows have nonzero entries */ \
69: PetscBool free_diag; \
70: datatype *a; /* nonzero elements */ \
71: PetscScalar *solve_work; /* work space used in MatSolve */ \
72: IS row, col, icol; /* index sets, used for reorderings */ \
73: PetscBool pivotinblocks; /* pivot inside factorization of each diagonal block */ \
74: Mat parent; /* set if this matrix was formed with MatDuplicate(...,MAT_SHARE_NONZERO_PATTERN,....); \
75: means that this shares some data structures with the parent including diag, ilen, imax, i, j */ \
76: Mat_SubSppt *submatis1; /* used by MatCreateSubMatrices_MPIXAIJ_Local */ \
77: Mat_SeqAIJOps ops[1] /* operations for SeqAIJ and its subclasses */
79: typedef struct {
80: MatTransposeColoring matcoloring;
81: Mat Bt_den; /* dense matrix of B^T */
82: Mat ABt_den; /* dense matrix of A*B^T */
83: PetscBool usecoloring;
84: } Mat_MatMatTransMult;
86: typedef struct { /* used by MatTransposeMatMult() */
87: Mat At; /* transpose of the first matrix */
88: Mat mA; /* maij matrix of A */
89: Vec bt, ct; /* vectors to hold locally transposed arrays of B and C */
90: /* used by PtAP */
91: void *data;
92: PetscErrorCode (*destroy)(void *);
93: } Mat_MatTransMatMult;
95: typedef struct {
96: PetscInt *api, *apj; /* symbolic structure of A*P */
97: PetscScalar *apa; /* temporary array for storing one row of A*P */
98: } Mat_AP;
100: typedef struct {
101: MatTransposeColoring matcoloring;
102: Mat Rt; /* sparse or dense matrix of R^T */
103: Mat RARt; /* dense matrix of R*A*R^T */
104: Mat ARt; /* A*R^T used for the case -matrart_color_art */
105: MatScalar *work; /* work array to store columns of A*R^T used in MatMatMatMultNumeric_SeqAIJ_SeqAIJ_SeqDense() */
106: /* free intermediate products needed for PtAP */
107: void *data;
108: PetscErrorCode (*destroy)(void *);
109: } Mat_RARt;
111: typedef struct {
112: Mat BC; /* temp matrix for storing B*C */
113: } Mat_MatMatMatMult;
115: /*
116: MATSEQAIJ format - Compressed row storage (also called Yale sparse matrix
117: format) or compressed sparse row (CSR). The i[] and j[] arrays start at 0. For example,
118: j[i[k]+p] is the pth column in row k. Note that the diagonal
119: matrix elements are stored with the rest of the nonzeros (not separately).
120: */
122: /* Info about i-nodes (identical nodes) helper class for SeqAIJ */
123: typedef struct {
124: MatScalar *bdiag, *ibdiag, *ssor_work; /* diagonal blocks of matrix used for MatSOR_SeqAIJ_Inode() */
125: PetscInt bdiagsize; /* length of bdiag and ibdiag */
126: PetscBool ibdiagvalid; /* do ibdiag[] and bdiag[] contain the most recent values */
128: PetscBool use;
129: PetscInt node_count; /* number of inodes */
130: PetscInt *size; /* size of each inode */
131: PetscInt limit; /* inode limit */
132: PetscInt max_limit; /* maximum supported inode limit */
133: PetscBool checked; /* if inodes have been checked for */
134: PetscObjectState mat_nonzerostate; /* non-zero state when inodes were checked for */
135: } Mat_SeqAIJ_Inode;
137: PETSC_INTERN PetscErrorCode MatView_SeqAIJ_Inode(Mat, PetscViewer);
138: PETSC_INTERN PetscErrorCode MatAssemblyEnd_SeqAIJ_Inode(Mat, MatAssemblyType);
139: PETSC_INTERN PetscErrorCode MatDestroy_SeqAIJ_Inode(Mat);
140: PETSC_INTERN PetscErrorCode MatCreate_SeqAIJ_Inode(Mat);
141: PETSC_INTERN PetscErrorCode MatSetOption_SeqAIJ_Inode(Mat, MatOption, PetscBool);
142: PETSC_INTERN PetscErrorCode MatDuplicate_SeqAIJ_Inode(Mat, MatDuplicateOption, Mat *);
143: PETSC_INTERN PetscErrorCode MatDuplicateNoCreate_SeqAIJ(Mat, Mat, MatDuplicateOption, PetscBool);
144: PETSC_INTERN PetscErrorCode MatLUFactorNumeric_SeqAIJ_Inode(Mat, Mat, const MatFactorInfo *);
145: PETSC_INTERN PetscErrorCode MatSeqAIJGetArray_SeqAIJ(Mat, PetscScalar **);
146: PETSC_INTERN PetscErrorCode MatSeqAIJRestoreArray_SeqAIJ(Mat, PetscScalar **);
148: typedef struct {
149: SEQAIJHEADER(MatScalar);
150: Mat_SeqAIJ_Inode inode;
151: MatScalar *saved_values; /* location for stashing nonzero values of matrix */
153: PetscScalar *idiag, *mdiag, *ssor_work; /* inverse of diagonal entries, diagonal values and workspace for Eisenstat trick */
154: PetscBool idiagvalid; /* current idiag[] and mdiag[] are valid */
155: PetscScalar *ibdiag; /* inverses of block diagonals */
156: PetscBool ibdiagvalid; /* inverses of block diagonals are valid. */
157: PetscBool diagonaldense; /* all entries along the diagonal have been set; i.e. no missing diagonal terms */
158: PetscScalar fshift, omega; /* last used omega and fshift */
160: /* MatSetValues() via hash related fields */
161: PetscHMapIJV ht;
162: PetscInt *dnz;
163: struct _MatOps cops;
164: } Mat_SeqAIJ;
166: typedef struct {
167: PetscInt nz; /* nz of the matrix after assembly */
168: PetscCount n; /* Number of entries in MatSetPreallocationCOO() */
169: PetscCount Atot; /* Total number of valid (i.e., w/ non-negative indices) entries in the COO array */
170: PetscCount *jmap; /* perm[jmap[i]..jmap[i+1]) give indices of entries in v[] associated with i-th nonzero of the matrix */
171: PetscCount *perm; /* The permutation array in sorting (i,j) by row and then by col */
172: } MatCOOStruct_SeqAIJ;
174: #define MatSeqXAIJGetOptions_Private(A) \
175: { \
176: const PetscBool oldvalues = (PetscBool)(A != PETSC_NULLPTR); \
177: PetscInt nonew = 0, nounused = 0; \
178: PetscBool roworiented = PETSC_FALSE; \
179: if (oldvalues) { \
180: nonew = ((Mat_SeqAIJ *)A->data)->nonew; \
181: nounused = ((Mat_SeqAIJ *)A->data)->nounused; \
182: roworiented = ((Mat_SeqAIJ *)A->data)->roworiented; \
183: } \
184: (void)0
186: #define MatSeqXAIJRestoreOptions_Private(A) \
187: if (oldvalues) { \
188: ((Mat_SeqAIJ *)A->data)->nonew = nonew; \
189: ((Mat_SeqAIJ *)A->data)->nounused = nounused; \
190: ((Mat_SeqAIJ *)A->data)->roworiented = roworiented; \
191: } \
192: } \
193: (void)0
195: /*
196: Frees the a, i, and j arrays from the XAIJ (AIJ, BAIJ, and SBAIJ) matrix types
197: */
198: static inline PetscErrorCode MatSeqXAIJFreeAIJ(Mat AA, MatScalar **a, PetscInt **j, PetscInt **i)
199: {
200: Mat_SeqAIJ *A = (Mat_SeqAIJ *)AA->data;
202: PetscFunctionBegin;
203: if (A->singlemalloc) {
204: PetscCall(PetscFree3(*a, *j, *i));
205: } else {
206: if (A->free_a) PetscCall(PetscFree(*a));
207: if (A->free_ij) PetscCall(PetscFree(*j));
208: if (A->free_ij) PetscCall(PetscFree(*i));
209: }
210: PetscFunctionReturn(PETSC_SUCCESS);
211: }
212: /*
213: Allocates larger a, i, and j arrays for the XAIJ (AIJ, BAIJ, and SBAIJ) matrix types
214: This is a macro because it takes the datatype as an argument which can be either a Mat or a MatScalar
215: */
216: #define MatSeqXAIJReallocateAIJ(Amat, AM, BS2, NROW, ROW, COL, RMAX, AA, AI, AJ, RP, AP, AIMAX, NONEW, datatype) \
217: do { \
218: if (NROW >= RMAX) { \
219: Mat_SeqAIJ *Ain = (Mat_SeqAIJ *)Amat->data; \
220: /* there is no extra room in row, therefore enlarge */ \
221: PetscInt CHUNKSIZE = 15, new_nz = AI[AM] + CHUNKSIZE, len, *new_i = NULL, *new_j = NULL; \
222: datatype *new_a; \
223: \
224: PetscCheck(NONEW != -2, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "New nonzero at (%" PetscInt_FMT ",%" PetscInt_FMT ") caused a malloc. Use MatSetOption(A, MAT_NEW_NONZERO_ALLOCATION_ERR, PETSC_FALSE) to turn off this check", ROW, COL); \
225: /* malloc new storage space */ \
226: PetscCall(PetscMalloc3(BS2 *new_nz, &new_a, new_nz, &new_j, AM + 1, &new_i)); \
227: \
228: /* copy over old data into new slots */ \
229: for (ii = 0; ii < ROW + 1; ii++) new_i[ii] = AI[ii]; \
230: for (ii = ROW + 1; ii < AM + 1; ii++) new_i[ii] = AI[ii] + CHUNKSIZE; \
231: PetscCall(PetscArraycpy(new_j, AJ, AI[ROW] + NROW)); \
232: len = (new_nz - CHUNKSIZE - AI[ROW] - NROW); \
233: PetscCall(PetscArraycpy(new_j + AI[ROW] + NROW + CHUNKSIZE, PetscSafePointerPlusOffset(AJ, AI[ROW] + NROW), len)); \
234: PetscCall(PetscArraycpy(new_a, AA, BS2 *(AI[ROW] + NROW))); \
235: PetscCall(PetscArrayzero(new_a + BS2 * (AI[ROW] + NROW), BS2 * CHUNKSIZE)); \
236: PetscCall(PetscArraycpy(new_a + BS2 * (AI[ROW] + NROW + CHUNKSIZE), PetscSafePointerPlusOffset(AA, BS2 * (AI[ROW] + NROW)), BS2 * len)); \
237: /* free up old matrix storage */ \
238: PetscCall(MatSeqXAIJFreeAIJ(A, &Ain->a, &Ain->j, &Ain->i)); \
239: AA = new_a; \
240: Ain->a = (MatScalar *)new_a; \
241: AI = Ain->i = new_i; \
242: AJ = Ain->j = new_j; \
243: Ain->singlemalloc = PETSC_TRUE; \
244: \
245: RP = AJ + AI[ROW]; \
246: AP = AA + BS2 * AI[ROW]; \
247: RMAX = AIMAX[ROW] = AIMAX[ROW] + CHUNKSIZE; \
248: Ain->maxnz += BS2 * CHUNKSIZE; \
249: Ain->reallocs++; \
250: } \
251: } while (0)
253: #define MatSeqXAIJReallocateAIJ_structure_only(Amat, AM, BS2, NROW, ROW, COL, RMAX, AI, AJ, RP, AIMAX, NONEW, datatype) \
254: do { \
255: if (NROW >= RMAX) { \
256: Mat_SeqAIJ *Ain = (Mat_SeqAIJ *)Amat->data; \
257: /* there is no extra room in row, therefore enlarge */ \
258: PetscInt CHUNKSIZE = 15, new_nz = AI[AM] + CHUNKSIZE, len, *new_i = NULL, *new_j = NULL; \
259: \
260: PetscCheck(NONEW != -2, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "New nonzero at (%" PetscInt_FMT ",%" PetscInt_FMT ") caused a malloc. Use MatSetOption(A, MAT_NEW_NONZERO_ALLOCATION_ERR, PETSC_FALSE) to turn off this check", ROW, COL); \
261: /* malloc new storage space */ \
262: PetscCall(PetscMalloc1(new_nz, &new_j)); \
263: PetscCall(PetscMalloc1(AM + 1, &new_i)); \
264: \
265: /* copy over old data into new slots */ \
266: for (ii = 0; ii < ROW + 1; ii++) new_i[ii] = AI[ii]; \
267: for (ii = ROW + 1; ii < AM + 1; ii++) new_i[ii] = AI[ii] + CHUNKSIZE; \
268: PetscCall(PetscArraycpy(new_j, AJ, AI[ROW] + NROW)); \
269: len = (new_nz - CHUNKSIZE - AI[ROW] - NROW); \
270: PetscCall(PetscArraycpy(new_j + AI[ROW] + NROW + CHUNKSIZE, AJ + AI[ROW] + NROW, len)); \
271: \
272: /* free up old matrix storage */ \
273: PetscCall(MatSeqXAIJFreeAIJ(A, &Ain->a, &Ain->j, &Ain->i)); \
274: Ain->a = NULL; \
275: AI = Ain->i = new_i; \
276: AJ = Ain->j = new_j; \
277: Ain->singlemalloc = PETSC_FALSE; \
278: Ain->free_a = PETSC_FALSE; \
279: \
280: RP = AJ + AI[ROW]; \
281: RMAX = AIMAX[ROW] = AIMAX[ROW] + CHUNKSIZE; \
282: Ain->maxnz += BS2 * CHUNKSIZE; \
283: Ain->reallocs++; \
284: } \
285: } while (0)
287: PETSC_INTERN PetscErrorCode MatSeqAIJSetPreallocation_SeqAIJ(Mat, PetscInt, const PetscInt *);
288: PETSC_INTERN PetscErrorCode MatSetPreallocationCOO_SeqAIJ(Mat, PetscCount, PetscInt[], PetscInt[]);
290: PETSC_INTERN PetscErrorCode MatILUFactorSymbolic_SeqAIJ(Mat, Mat, IS, IS, const MatFactorInfo *);
291: PETSC_INTERN PetscErrorCode MatILUFactorSymbolic_SeqAIJ_ilu0(Mat, Mat, IS, IS, const MatFactorInfo *);
293: PETSC_INTERN PetscErrorCode MatICCFactorSymbolic_SeqAIJ(Mat, Mat, IS, const MatFactorInfo *);
294: PETSC_INTERN PetscErrorCode MatCholeskyFactorSymbolic_SeqAIJ(Mat, Mat, IS, const MatFactorInfo *);
295: PETSC_INTERN PetscErrorCode MatCholeskyFactorNumeric_SeqAIJ_inplace(Mat, Mat, const MatFactorInfo *);
296: PETSC_INTERN PetscErrorCode MatCholeskyFactorNumeric_SeqAIJ(Mat, Mat, const MatFactorInfo *);
297: PETSC_INTERN PetscErrorCode MatDuplicate_SeqAIJ(Mat, MatDuplicateOption, Mat *);
298: PETSC_INTERN PetscErrorCode MatCopy_SeqAIJ(Mat, Mat, MatStructure);
299: PETSC_INTERN PetscErrorCode MatMissingDiagonal_SeqAIJ(Mat, PetscBool *, PetscInt *);
300: PETSC_INTERN PetscErrorCode MatMarkDiagonal_SeqAIJ(Mat);
301: PETSC_INTERN PetscErrorCode MatFindZeroDiagonals_SeqAIJ_Private(Mat, PetscInt *, PetscInt **);
303: PETSC_INTERN PetscErrorCode MatMult_SeqAIJ(Mat, Vec, Vec);
304: PETSC_INTERN PetscErrorCode MatMult_SeqAIJ_Inode(Mat, Vec, Vec);
305: PETSC_INTERN PetscErrorCode MatMultAdd_SeqAIJ(Mat, Vec, Vec, Vec);
306: PETSC_INTERN PetscErrorCode MatMultAdd_SeqAIJ_Inode(Mat, Vec, Vec, Vec);
307: PETSC_INTERN PetscErrorCode MatMultTranspose_SeqAIJ(Mat, Vec, Vec);
308: PETSC_INTERN PetscErrorCode MatMultTransposeAdd_SeqAIJ(Mat, Vec, Vec, Vec);
309: PETSC_INTERN PetscErrorCode MatSOR_SeqAIJ(Mat, Vec, PetscReal, MatSORType, PetscReal, PetscInt, PetscInt, Vec);
310: PETSC_INTERN PetscErrorCode MatSOR_SeqAIJ_Inode(Mat, Vec, PetscReal, MatSORType, PetscReal, PetscInt, PetscInt, Vec);
312: PETSC_INTERN PetscErrorCode MatSetOption_SeqAIJ(Mat, MatOption, PetscBool);
314: PETSC_INTERN PetscErrorCode MatGetSymbolicTranspose_SeqAIJ(Mat, PetscInt *[], PetscInt *[]);
315: PETSC_INTERN PetscErrorCode MatRestoreSymbolicTranspose_SeqAIJ(Mat, PetscInt *[], PetscInt *[]);
316: PETSC_INTERN PetscErrorCode MatGetSymbolicTransposeReduced_SeqAIJ(Mat, PetscInt, PetscInt, PetscInt *[], PetscInt *[]);
317: PETSC_INTERN PetscErrorCode MatTransposeSymbolic_SeqAIJ(Mat, Mat *);
318: PETSC_INTERN PetscErrorCode MatTranspose_SeqAIJ(Mat, MatReuse, Mat *);
320: PETSC_INTERN PetscErrorCode MatToSymmetricIJ_SeqAIJ(PetscInt, PetscInt *, PetscInt *, PetscBool, PetscInt, PetscInt, PetscInt **, PetscInt **);
321: PETSC_INTERN PetscErrorCode MatLUFactorSymbolic_SeqAIJ(Mat, Mat, IS, IS, const MatFactorInfo *);
322: PETSC_INTERN PetscErrorCode MatLUFactorNumeric_SeqAIJ_inplace(Mat, Mat, const MatFactorInfo *);
323: PETSC_INTERN PetscErrorCode MatLUFactorNumeric_SeqAIJ(Mat, Mat, const MatFactorInfo *);
324: PETSC_INTERN PetscErrorCode MatLUFactorNumeric_SeqAIJ_InplaceWithPerm(Mat, Mat, const MatFactorInfo *);
325: PETSC_INTERN PetscErrorCode MatLUFactor_SeqAIJ(Mat, IS, IS, const MatFactorInfo *);
326: PETSC_INTERN PetscErrorCode MatSolve_SeqAIJ_inplace(Mat, Vec, Vec);
327: PETSC_INTERN PetscErrorCode MatSolve_SeqAIJ(Mat, Vec, Vec);
328: PETSC_INTERN PetscErrorCode MatSolve_SeqAIJ_Inode(Mat, Vec, Vec);
329: PETSC_INTERN PetscErrorCode MatSolve_SeqAIJ_NaturalOrdering(Mat, Vec, Vec);
330: PETSC_INTERN PetscErrorCode MatSolveAdd_SeqAIJ(Mat, Vec, Vec, Vec);
331: PETSC_INTERN PetscErrorCode MatSolveTranspose_SeqAIJ_inplace(Mat, Vec, Vec);
332: PETSC_INTERN PetscErrorCode MatSolveTranspose_SeqAIJ(Mat, Vec, Vec);
333: PETSC_INTERN PetscErrorCode MatSolveTransposeAdd_SeqAIJ_inplace(Mat, Vec, Vec, Vec);
334: PETSC_INTERN PetscErrorCode MatSolveTransposeAdd_SeqAIJ(Mat, Vec, Vec, Vec);
335: PETSC_INTERN PetscErrorCode MatMatSolve_SeqAIJ(Mat, Mat, Mat);
336: PETSC_INTERN PetscErrorCode MatMatSolveTranspose_SeqAIJ(Mat, Mat, Mat);
337: PETSC_INTERN PetscErrorCode MatEqual_SeqAIJ(Mat, Mat, PetscBool *);
338: PETSC_INTERN PetscErrorCode MatFDColoringCreate_SeqXAIJ(Mat, ISColoring, MatFDColoring);
339: PETSC_INTERN PetscErrorCode MatFDColoringSetUp_SeqXAIJ(Mat, ISColoring, MatFDColoring);
340: PETSC_INTERN PetscErrorCode MatFDColoringSetUpBlocked_AIJ_Private(Mat, MatFDColoring, PetscInt);
341: PETSC_INTERN PetscErrorCode MatLoad_AIJ_HDF5(Mat, PetscViewer);
342: PETSC_INTERN PetscErrorCode MatLoad_SeqAIJ_Binary(Mat, PetscViewer);
343: PETSC_INTERN PetscErrorCode MatLoad_SeqAIJ(Mat, PetscViewer);
344: PETSC_INTERN PetscErrorCode RegisterApplyPtAPRoutines_Private(Mat);
346: #if defined(PETSC_HAVE_HYPRE)
347: PETSC_INTERN PetscErrorCode MatProductSetFromOptions_Transpose_AIJ_AIJ(Mat);
348: #endif
349: PETSC_INTERN PetscErrorCode MatProductSetFromOptions_SeqAIJ(Mat);
351: PETSC_INTERN PetscErrorCode MatProductSymbolic_SeqAIJ_SeqAIJ(Mat);
352: PETSC_INTERN PetscErrorCode MatProductSymbolic_PtAP_SeqAIJ_SeqAIJ(Mat);
353: PETSC_INTERN PetscErrorCode MatProductSymbolic_RARt_SeqAIJ_SeqAIJ(Mat);
355: PETSC_INTERN PetscErrorCode MatMatMultSymbolic_SeqAIJ_SeqAIJ(Mat, Mat, PetscReal, Mat);
356: PETSC_INTERN PetscErrorCode MatMatMultSymbolic_SeqAIJ_SeqAIJ_Sorted(Mat, Mat, PetscReal, Mat);
357: PETSC_INTERN PetscErrorCode MatMatMultSymbolic_SeqDense_SeqAIJ(Mat, Mat, PetscReal, Mat);
358: PETSC_INTERN PetscErrorCode MatMatMultSymbolic_SeqAIJ_SeqAIJ_Scalable(Mat, Mat, PetscReal, Mat);
359: PETSC_INTERN PetscErrorCode MatMatMultSymbolic_SeqAIJ_SeqAIJ_Scalable_fast(Mat, Mat, PetscReal, Mat);
360: PETSC_INTERN PetscErrorCode MatMatMultSymbolic_SeqAIJ_SeqAIJ_Heap(Mat, Mat, PetscReal, Mat);
361: PETSC_INTERN PetscErrorCode MatMatMultSymbolic_SeqAIJ_SeqAIJ_BTHeap(Mat, Mat, PetscReal, Mat);
362: PETSC_INTERN PetscErrorCode MatMatMultSymbolic_SeqAIJ_SeqAIJ_RowMerge(Mat, Mat, PetscReal, Mat);
363: PETSC_INTERN PetscErrorCode MatMatMultSymbolic_SeqAIJ_SeqAIJ_LLCondensed(Mat, Mat, PetscReal, Mat);
364: #if defined(PETSC_HAVE_HYPRE)
365: PETSC_INTERN PetscErrorCode MatMatMultSymbolic_AIJ_AIJ_wHYPRE(Mat, Mat, PetscReal, Mat);
366: #endif
368: PETSC_INTERN PetscErrorCode MatMatMultNumeric_SeqAIJ_SeqAIJ(Mat, Mat, Mat);
369: PETSC_INTERN PetscErrorCode MatMatMultNumeric_SeqAIJ_SeqAIJ_Sorted(Mat, Mat, Mat);
371: PETSC_INTERN PetscErrorCode MatMatMultNumeric_SeqDense_SeqAIJ(Mat, Mat, Mat);
372: PETSC_INTERN PetscErrorCode MatMatMultNumeric_SeqAIJ_SeqAIJ_Scalable(Mat, Mat, Mat);
374: PETSC_INTERN PetscErrorCode MatPtAPSymbolic_SeqAIJ_SeqAIJ_SparseAxpy(Mat, Mat, PetscReal, Mat);
375: PETSC_INTERN PetscErrorCode MatPtAPNumeric_SeqAIJ_SeqAIJ(Mat, Mat, Mat);
376: PETSC_INTERN PetscErrorCode MatPtAPNumeric_SeqAIJ_SeqAIJ_SparseAxpy(Mat, Mat, Mat);
378: PETSC_INTERN PetscErrorCode MatRARtSymbolic_SeqAIJ_SeqAIJ(Mat, Mat, PetscReal, Mat);
379: PETSC_INTERN PetscErrorCode MatRARtSymbolic_SeqAIJ_SeqAIJ_matmattransposemult(Mat, Mat, PetscReal, Mat);
380: PETSC_INTERN PetscErrorCode MatRARtSymbolic_SeqAIJ_SeqAIJ_colorrart(Mat, Mat, PetscReal, Mat);
381: PETSC_INTERN PetscErrorCode MatRARtNumeric_SeqAIJ_SeqAIJ(Mat, Mat, Mat);
382: PETSC_INTERN PetscErrorCode MatRARtNumeric_SeqAIJ_SeqAIJ_matmattransposemult(Mat, Mat, Mat);
383: PETSC_INTERN PetscErrorCode MatRARtNumeric_SeqAIJ_SeqAIJ_colorrart(Mat, Mat, Mat);
385: PETSC_INTERN PetscErrorCode MatTransposeMatMultSymbolic_SeqAIJ_SeqAIJ(Mat, Mat, PetscReal, Mat);
386: PETSC_INTERN PetscErrorCode MatTransposeMatMultNumeric_SeqAIJ_SeqAIJ(Mat, Mat, Mat);
387: PETSC_INTERN PetscErrorCode MatDestroy_SeqAIJ_MatTransMatMult(void *);
389: PETSC_INTERN PetscErrorCode MatMatTransposeMultSymbolic_SeqAIJ_SeqAIJ(Mat, Mat, PetscReal, Mat);
390: PETSC_INTERN PetscErrorCode MatMatTransposeMultNumeric_SeqAIJ_SeqAIJ(Mat, Mat, Mat);
391: PETSC_INTERN PetscErrorCode MatTransposeColoringCreate_SeqAIJ(Mat, ISColoring, MatTransposeColoring);
392: PETSC_INTERN PetscErrorCode MatTransColoringApplySpToDen_SeqAIJ(MatTransposeColoring, Mat, Mat);
393: PETSC_INTERN PetscErrorCode MatTransColoringApplyDenToSp_SeqAIJ(MatTransposeColoring, Mat, Mat);
395: PETSC_INTERN PetscErrorCode MatMatMatMultSymbolic_SeqAIJ_SeqAIJ_SeqAIJ(Mat, Mat, Mat, PetscReal, Mat);
396: PETSC_INTERN PetscErrorCode MatMatMatMultNumeric_SeqAIJ_SeqAIJ_SeqAIJ(Mat, Mat, Mat, Mat);
398: PETSC_INTERN PetscErrorCode MatSetRandomSkipColumnRange_SeqAIJ_Private(Mat, PetscInt, PetscInt, PetscRandom);
399: PETSC_INTERN PetscErrorCode MatSetValues_SeqAIJ(Mat, PetscInt, const PetscInt[], PetscInt, const PetscInt[], const PetscScalar[], InsertMode);
400: PETSC_INTERN PetscErrorCode MatGetRow_SeqAIJ(Mat, PetscInt, PetscInt *, PetscInt **, PetscScalar **);
401: PETSC_INTERN PetscErrorCode MatRestoreRow_SeqAIJ(Mat, PetscInt, PetscInt *, PetscInt **, PetscScalar **);
402: PETSC_INTERN PetscErrorCode MatScale_SeqAIJ(Mat, PetscScalar);
403: PETSC_INTERN PetscErrorCode MatDiagonalScale_SeqAIJ(Mat, Vec, Vec);
404: PETSC_INTERN PetscErrorCode MatDiagonalSet_SeqAIJ(Mat, Vec, InsertMode);
405: PETSC_INTERN PetscErrorCode MatAXPY_SeqAIJ(Mat, PetscScalar, Mat, MatStructure);
406: PETSC_INTERN PetscErrorCode MatGetRowIJ_SeqAIJ(Mat, PetscInt, PetscBool, PetscBool, PetscInt *, const PetscInt *[], const PetscInt *[], PetscBool *);
407: PETSC_INTERN PetscErrorCode MatRestoreRowIJ_SeqAIJ(Mat, PetscInt, PetscBool, PetscBool, PetscInt *, const PetscInt *[], const PetscInt *[], PetscBool *);
408: PETSC_INTERN PetscErrorCode MatGetColumnIJ_SeqAIJ(Mat, PetscInt, PetscBool, PetscBool, PetscInt *, const PetscInt *[], const PetscInt *[], PetscBool *);
409: PETSC_INTERN PetscErrorCode MatRestoreColumnIJ_SeqAIJ(Mat, PetscInt, PetscBool, PetscBool, PetscInt *, const PetscInt *[], const PetscInt *[], PetscBool *);
410: PETSC_INTERN PetscErrorCode MatGetColumnIJ_SeqAIJ_Color(Mat, PetscInt, PetscBool, PetscBool, PetscInt *, const PetscInt *[], const PetscInt *[], PetscInt *[], PetscBool *);
411: PETSC_INTERN PetscErrorCode MatRestoreColumnIJ_SeqAIJ_Color(Mat, PetscInt, PetscBool, PetscBool, PetscInt *, const PetscInt *[], const PetscInt *[], PetscInt *[], PetscBool *);
412: PETSC_INTERN PetscErrorCode MatDestroy_SeqAIJ(Mat);
413: PETSC_INTERN PetscErrorCode MatView_SeqAIJ(Mat, PetscViewer);
415: PETSC_INTERN PetscErrorCode MatSeqAIJInvalidateDiagonal(Mat);
416: PETSC_INTERN PetscErrorCode MatSeqAIJInvalidateDiagonal_Inode(Mat);
417: PETSC_INTERN PetscErrorCode MatSeqAIJCheckInode(Mat);
418: PETSC_INTERN PetscErrorCode MatSeqAIJCheckInode_FactorLU(Mat);
420: PETSC_INTERN PetscErrorCode MatAXPYGetPreallocation_SeqAIJ(Mat, Mat, PetscInt *);
422: #if defined(PETSC_HAVE_MATLAB)
423: PETSC_EXTERN PetscErrorCode MatlabEnginePut_SeqAIJ(PetscObject, void *);
424: PETSC_EXTERN PetscErrorCode MatlabEngineGet_SeqAIJ(PetscObject, void *);
425: #endif
426: PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqSBAIJ(Mat, MatType, MatReuse, Mat *);
427: PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqBAIJ(Mat, MatType, MatReuse, Mat *);
428: PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqDense(Mat, MatType, MatReuse, Mat *);
429: PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqAIJCRL(Mat, MatType, MatReuse, Mat *);
430: PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_Elemental(Mat, MatType, MatReuse, Mat *);
431: #if defined(PETSC_HAVE_SCALAPACK)
432: PETSC_INTERN PetscErrorCode MatConvert_AIJ_ScaLAPACK(Mat, MatType, MatReuse, Mat *);
433: #endif
434: PETSC_INTERN PetscErrorCode MatConvert_AIJ_HYPRE(Mat, MatType, MatReuse, Mat *);
435: PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqAIJPERM(Mat, MatType, MatReuse, Mat *);
436: PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqAIJSELL(Mat, MatType, MatReuse, Mat *);
437: PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqAIJMKL(Mat, MatType, MatReuse, Mat *);
438: PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqAIJViennaCL(Mat, MatType, MatReuse, Mat *);
439: PETSC_INTERN PetscErrorCode MatReorderForNonzeroDiagonal_SeqAIJ(Mat, PetscReal, IS, IS);
440: PETSC_INTERN PetscErrorCode MatRARt_SeqAIJ_SeqAIJ(Mat, Mat, MatReuse, PetscReal, Mat *);
441: PETSC_EXTERN PetscErrorCode MatCreate_SeqAIJ(Mat);
442: PETSC_INTERN PetscErrorCode MatAssemblyEnd_SeqAIJ(Mat, MatAssemblyType);
443: PETSC_EXTERN PetscErrorCode MatZeroEntries_SeqAIJ(Mat);
445: PETSC_INTERN PetscErrorCode MatAXPYGetPreallocation_SeqX_private(PetscInt, const PetscInt *, const PetscInt *, const PetscInt *, const PetscInt *, PetscInt *);
446: PETSC_INTERN PetscErrorCode MatCreateMPIMatConcatenateSeqMat_SeqAIJ(MPI_Comm, Mat, PetscInt, MatReuse, Mat *);
447: PETSC_INTERN PetscErrorCode MatCreateMPIMatConcatenateSeqMat_MPIAIJ(MPI_Comm, Mat, PetscInt, MatReuse, Mat *);
449: PETSC_INTERN PetscErrorCode MatSetSeqMat_SeqAIJ(Mat, IS, IS, MatStructure, Mat);
450: PETSC_INTERN PetscErrorCode MatEliminateZeros_SeqAIJ(Mat, PetscBool);
451: PETSC_INTERN PetscErrorCode MatDestroySubMatrix_Private(Mat_SubSppt *);
452: PETSC_INTERN PetscErrorCode MatDestroySubMatrix_SeqAIJ(Mat);
453: PETSC_INTERN PetscErrorCode MatDestroySubMatrix_Dummy(Mat);
454: PETSC_INTERN PetscErrorCode MatDestroySubMatrices_Dummy(PetscInt, Mat *[]);
455: PETSC_INTERN PetscErrorCode MatCreateSubMatrix_SeqAIJ(Mat, IS, IS, PetscInt, MatReuse, Mat *);
457: PETSC_INTERN PetscErrorCode MatSetSeqAIJWithArrays_private(MPI_Comm, PetscInt, PetscInt, PetscInt[], PetscInt[], PetscScalar[], MatType, Mat);
459: PETSC_SINGLE_LIBRARY_INTERN PetscErrorCode MatSeqAIJCompactOutExtraColumns_SeqAIJ(Mat, ISLocalToGlobalMapping *);
461: /*
462: PetscSparseDenseMinusDot - The inner kernel of triangular solves and Gauss-Siedel smoothing. \sum_i xv[i] * r[xi[i]] for CSR storage
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
470: Output Parameter:
471: . sum - negative the sum of results
473: PETSc compile flags:
474: + PETSC_KERNEL_USE_UNROLL_4
475: - PETSC_KERNEL_USE_UNROLL_2
477: Developer Note:
478: The macro changes sum but not other parameters
480: .seealso: `PetscSparseDensePlusDot()`
481: */
482: #if defined(PETSC_KERNEL_USE_UNROLL_4)
483: #define PetscSparseDenseMinusDot(sum, r, xv, xi, nnz) \
484: do { \
485: if (nnz > 0) { \
486: PetscInt nnz2 = nnz, rem = nnz & 0x3; \
487: switch (rem) { \
488: case 3: \
489: sum -= *xv++ * r[*xi++]; \
490: case 2: \
491: sum -= *xv++ * r[*xi++]; \
492: case 1: \
493: sum -= *xv++ * r[*xi++]; \
494: nnz2 -= rem; \
495: } \
496: while (nnz2 > 0) { \
497: sum -= xv[0] * r[xi[0]] + xv[1] * r[xi[1]] + xv[2] * r[xi[2]] + xv[3] * r[xi[3]]; \
498: xv += 4; \
499: xi += 4; \
500: nnz2 -= 4; \
501: } \
502: xv -= nnz; \
503: xi -= nnz; \
504: } \
505: } while (0)
507: #elif defined(PETSC_KERNEL_USE_UNROLL_2)
508: #define PetscSparseDenseMinusDot(sum, r, xv, xi, nnz) \
509: do { \
510: PetscInt __i, __i1, __i2; \
511: for (__i = 0; __i < nnz - 1; __i += 2) { \
512: __i1 = xi[__i]; \
513: __i2 = xi[__i + 1]; \
514: sum -= (xv[__i] * r[__i1] + xv[__i + 1] * r[__i2]); \
515: } \
516: if (nnz & 0x1) sum -= xv[__i] * r[xi[__i]]; \
517: } while (0)
519: #else
520: #define PetscSparseDenseMinusDot(sum, r, xv, xi, nnz) \
521: do { \
522: PetscInt __i; \
523: for (__i = 0; __i < nnz; __i++) sum -= xv[__i] * r[xi[__i]]; \
524: } while (0)
525: #endif
527: /*
528: PetscSparseDensePlusDot - The inner kernel of matrix-vector product \sum_i xv[i] * r[xi[i]] for CSR storage
530: Input Parameters:
531: + nnz - the number of entries
532: . r - the array of vector values
533: . xv - the matrix values for the row
534: - xi - the column indices of the nonzeros in the row
536: Output Parameter:
537: . sum - the sum of results
539: PETSc compile flags:
540: + PETSC_KERNEL_USE_UNROLL_4
541: - PETSC_KERNEL_USE_UNROLL_2
543: Developer Note:
544: The macro changes sum but not other parameters
546: .seealso: `PetscSparseDenseMinusDot()`
547: */
548: #if defined(PETSC_KERNEL_USE_UNROLL_4)
549: #define PetscSparseDensePlusDot(sum, r, xv, xi, nnz) \
550: do { \
551: if (nnz > 0) { \
552: PetscInt nnz2 = nnz, rem = nnz & 0x3; \
553: switch (rem) { \
554: case 3: \
555: sum += *xv++ * r[*xi++]; \
556: case 2: \
557: sum += *xv++ * r[*xi++]; \
558: case 1: \
559: sum += *xv++ * r[*xi++]; \
560: nnz2 -= rem; \
561: } \
562: while (nnz2 > 0) { \
563: sum += xv[0] * r[xi[0]] + xv[1] * r[xi[1]] + xv[2] * r[xi[2]] + xv[3] * r[xi[3]]; \
564: xv += 4; \
565: xi += 4; \
566: nnz2 -= 4; \
567: } \
568: xv -= nnz; \
569: xi -= nnz; \
570: } \
571: } while (0)
573: #elif defined(PETSC_KERNEL_USE_UNROLL_2)
574: #define PetscSparseDensePlusDot(sum, r, xv, xi, nnz) \
575: do { \
576: PetscInt __i, __i1, __i2; \
577: for (__i = 0; __i < nnz - 1; __i += 2) { \
578: __i1 = xi[__i]; \
579: __i2 = xi[__i + 1]; \
580: sum += (xv[__i] * r[__i1] + xv[__i + 1] * r[__i2]); \
581: } \
582: if (nnz & 0x1) sum += xv[__i] * r[xi[__i]]; \
583: } while (0)
585: #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)
586: #define PetscSparseDensePlusDot(sum, r, xv, xi, nnz) PetscSparseDensePlusDot_AVX512_Private(&(sum), (r), (xv), (xi), (nnz))
588: #else
589: #define PetscSparseDensePlusDot(sum, r, xv, xi, nnz) \
590: do { \
591: PetscInt __i; \
592: for (__i = 0; __i < nnz; __i++) sum += xv[__i] * r[xi[__i]]; \
593: } while (0)
594: #endif
596: #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)
597: #include <immintrin.h>
598: #if !defined(_MM_SCALE_8)
599: #define _MM_SCALE_8 8
600: #endif
602: static inline void PetscSparseDensePlusDot_AVX512_Private(PetscScalar *sum, const PetscScalar *x, const MatScalar *aa, const PetscInt *aj, PetscInt n)
603: {
604: __m512d vec_x, vec_y, vec_vals;
605: __m256i vec_idx;
606: PetscInt j;
608: vec_y = _mm512_setzero_pd();
609: for (j = 0; j < (n >> 3); j++) {
610: vec_idx = _mm256_loadu_si256((__m256i const *)aj);
611: vec_vals = _mm512_loadu_pd(aa);
612: vec_x = _mm512_i32gather_pd(vec_idx, x, _MM_SCALE_8);
613: vec_y = _mm512_fmadd_pd(vec_x, vec_vals, vec_y);
614: aj += 8;
615: aa += 8;
616: }
617: #if defined(__AVX512VL__)
618: /* masked load requires avx512vl, which is not supported by KNL */
619: if (n & 0x07) {
620: __mmask8 mask;
621: mask = (__mmask8)(0xff >> (8 - (n & 0x07)));
622: vec_idx = _mm256_mask_loadu_epi32(vec_idx, mask, aj);
623: vec_vals = _mm512_mask_loadu_pd(vec_vals, mask, aa);
624: vec_x = _mm512_mask_i32gather_pd(vec_x, mask, vec_idx, x, _MM_SCALE_8);
625: vec_y = _mm512_mask3_fmadd_pd(vec_x, vec_vals, vec_y, mask);
626: }
627: *sum += _mm512_reduce_add_pd(vec_y);
628: #else
629: *sum += _mm512_reduce_add_pd(vec_y);
630: for (j = 0; j < (n & 0x07); j++) *sum += aa[j] * x[aj[j]];
631: #endif
632: }
633: #endif
635: /*
636: PetscSparseDenseMaxDot - The inner kernel of a modified matrix-vector product \max_i xv[i] * r[xi[i]] for CSR storage
638: Input Parameters:
639: + nnz - the number of entries
640: . r - the array of vector values
641: . xv - the matrix values for the row
642: - xi - the column indices of the nonzeros in the row
644: Output Parameter:
645: . max - the max of results
647: .seealso: `PetscSparseDensePlusDot()`, `PetscSparseDenseMinusDot()`
648: */
649: #define PetscSparseDenseMaxDot(max, r, xv, xi, nnz) \
650: do { \
651: for (PetscInt __i = 0; __i < (nnz); __i++) { max = PetscMax(PetscRealPart(max), PetscRealPart((xv)[__i] * (r)[(xi)[__i]])); } \
652: } while (0)
654: /*
655: Add column indices into table for counting the max nonzeros of merged rows
656: */
657: #define MatRowMergeMax_SeqAIJ(mat, nrows, ta) \
658: do { \
659: if (mat) { \
660: for (PetscInt _row = 0; _row < (nrows); _row++) { \
661: const PetscInt _nz = (mat)->i[_row + 1] - (mat)->i[_row]; \
662: for (PetscInt _j = 0; _j < _nz; _j++) { \
663: PetscInt *_col = _j + (mat)->j + (mat)->i[_row]; \
664: PetscCall(PetscHMapISet((ta), *_col + 1, 1)); \
665: } \
666: } \
667: } \
668: } while (0)
670: /*
671: Add column indices into table for counting the nonzeros of merged rows
672: */
673: #define MatMergeRows_SeqAIJ(mat, nrows, rows, ta) \
674: do { \
675: for (PetscInt _i = 0; _i < (nrows); _i++) { \
676: const PetscInt _row = (rows)[_i]; \
677: const PetscInt _nz = (mat)->i[_row + 1] - (mat)->i[_row]; \
678: for (PetscInt _j = 0; _j < _nz; _j++) { \
679: PetscInt *_col = _j + (mat)->j + (mat)->i[_row]; \
680: PetscCall(PetscHMapISetWithMode((ta), *_col + 1, 1, INSERT_VALUES)); \
681: } \
682: } \
683: } while (0)