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: PetscMPIInt nrqs, nrqr;
13: PetscInt **rbuf1, **rbuf2, **rbuf3, **sbuf1, **sbuf2;
14: PetscInt **ptr;
15: PetscInt *tmp;
16: PetscInt *ctr;
17: PetscMPIInt *pa; /* process array */
18: PetscInt *req_size;
19: PetscMPIInt *req_source1, *req_source2;
20: PetscBool allcolumns, allrows;
21: PetscBool singleis;
22: PetscBool csrcached; /* owned-row maps have been built, including a valid empty cache */
23: PetscMPIInt *row2proc; /* row to process (MPI rank) map */
24: PetscInt nlocal_a, nlocal_b; /* cached entries from the parent diagonal and off-diagonal blocks */
25: PetscInt *local_a_parent, *local_a_sub; /* positions in the parent diagonal and submatrix value arrays */
26: PetscInt *local_b_parent, *local_b_sub; /* positions in the parent off-diagonal and submatrix value arrays */
27: PetscInt nstages;
28: #if PetscDefined(USE_CTABLE)
29: PetscHMapI cmap, rmap;
30: PetscInt *cmap_loc, *rmap_loc;
31: #else
32: PetscInt *cmap, *rmap;
33: #endif
34: PetscObjectState nonzerostate; /* initial submatrix graph state; cached positions require it unchanged */
35: PetscErrorCode (*destroy)(Mat);
36: } Mat_SubSppt;
38: /* Operations provided by MATSEQAIJ and its subclasses */
39: typedef struct {
40: PetscErrorCode (*getarray)(Mat, PetscScalar **);
41: PetscErrorCode (*restorearray)(Mat, PetscScalar **);
42: PetscErrorCode (*getarrayread)(Mat, const PetscScalar **);
43: PetscErrorCode (*restorearrayread)(Mat, const PetscScalar **);
44: PetscErrorCode (*getarraywrite)(Mat, PetscScalar **);
45: PetscErrorCode (*restorearraywrite)(Mat, PetscScalar **);
46: PetscErrorCode (*getcsrandmemtype)(Mat, const PetscInt **, const PetscInt **, PetscScalar **, PetscMemType *);
47: } Mat_SeqAIJOps;
49: /*
50: Struct header shared by SeqAIJ, SeqBAIJ, and SeqSBAIJ matrix formats
51: */
52: #define SEQAIJHEADER(datatype) \
53: PetscBool roworiented; /* if true, row-oriented input, default */ \
54: PetscInt nonew; /* 1 don't add new nonzeros, -1 generate error on new */ \
55: PetscInt nounused; /* -1 generate error on unused space */ \
56: PetscInt maxnz; /* allocated nonzeros */ \
57: PetscInt *imax; /* maximum space allocated for each row */ \
58: PetscInt *ilen; /* actual length of each row */ \
59: PetscInt *ipre; /* space preallocated for each row by user */ \
60: PetscBool free_imax_ilen; \
61: PetscInt reallocs; /* number of mallocs done during MatSetValues() \
62: as more values are set than were prealloced */ \
63: PetscInt rmax; /* max nonzeros in any row */ \
64: PetscBool keepnonzeropattern; /* keeps matrix nonzero structure same in calls to MatZeroRows()*/ \
65: PetscBool ignorezeroentries; \
66: PetscBool free_ij; /* free the column indices j and row offsets i when the matrix is destroyed */ \
67: PetscBool free_a; /* free the numerical values when matrix is destroy */ \
68: Mat_CompressedRow compressedrow; /* use compressed row format */ \
69: PetscInt nz; /* nonzeros */ \
70: PetscInt *i; /* pointer to beginning of each row */ \
71: PetscInt *j; /* column values: j + i[k] - 1 is start of row k */ \
72: PetscInt *diag; /* pointers to diagonal elements */ \
73: PetscObjectState diagNonzeroState; /* nonzero state of the matrix when diag was obtained */ \
74: PetscBool diagDense; /* all entries along the diagonal have been set; i.e. no missing diagonal terms */ \
75: PetscInt nonzerorowcnt; /* how many rows have nonzero entries */ \
76: datatype *a; /* nonzero elements */ \
77: PetscScalar *solve_work; /* work space used in MatSolve */ \
78: IS row, col, icol; /* index sets, used for reorderings */ \
79: PetscBool pivotinblocks; /* pivot inside factorization of each diagonal block */ \
80: Mat parent; /* set if this matrix was formed with MatDuplicate(...,MAT_SHARE_NONZERO_PATTERN,....); \
81: means that this shares some data structures with the parent including diag, ilen, imax, i, j */ \
82: Mat_SubSppt *submatis1; /* used by MatCreateSubMatrices_MPIXAIJ_Local */ \
83: Mat_SeqAIJOps ops[1] /* operations for SeqAIJ and its subclasses */
85: typedef struct {
86: MatTransposeColoring matcoloring;
87: Mat Bt_den; /* dense matrix of B^T */
88: Mat ABt_den; /* dense matrix of A*B^T */
89: PetscBool usecoloring;
90: } MatProductCtx_MatMatTransMult;
92: typedef struct { /* used by MatTransposeMatMult() */
93: Mat At; /* transpose of the first matrix */
94: Mat mA; /* maij matrix of A */
95: Vec bt, ct; /* vectors to hold locally transposed arrays of B and C */
96: /* used by PtAP */
97: void *data;
98: PetscCtxDestroyFn *destroy;
99: } MatProductCtx_MatTransMatMult;
101: typedef struct {
102: PetscInt *api, *apj; /* symbolic structure of A*P */
103: PetscScalar *apa; /* temporary array for storing one row of A*P */
104: } MatProductCtx_AP;
106: typedef struct {
107: MatTransposeColoring matcoloring;
108: Mat Rt; /* sparse or dense matrix of R^T */
109: Mat RARt; /* dense matrix of R*A*R^T */
110: Mat ARt; /* A*R^T used for the case -matrart_color_art */
111: MatScalar *work; /* work array to store columns of A*R^T used in MatMatMatMultNumeric_SeqAIJ_SeqAIJ_SeqDense() */
112: /* free intermediate products needed for PtAP */
113: void *data;
114: PetscCtxDestroyFn *destroy;
115: } MatProductCtx_RARt;
117: typedef struct {
118: Mat BC; /* temp matrix for storing B*C */
119: } MatProductCtx_MatMatMatMult;
121: /*
122: MATSEQAIJ format - Compressed row storage (also called Yale sparse matrix
123: format) or compressed sparse row (CSR). The i[] and j[] arrays start at 0. For example,
124: j[i[k]+p] is the pth column in row k. Note that the diagonal
125: matrix elements are stored with the rest of the nonzeros (not separately).
126: */
128: /* Info about i-nodes (identical nodes) helper class for SeqAIJ */
129: typedef struct {
130: /* data for MatSOR_SeqAIJ_Inode() */
131: MatScalar *bdiag, *ibdiag, *ssor_work; /* diagonal blocks of matrices */
132: PetscInt bdiagsize; /* length of bdiag and ibdiag */
133: PetscObjectState ibdiagState; /* state of the matrix when ibdiag[] and bdiag[] were constructed */
135: PetscBool use;
136: PetscInt node_count; /* number of inodes */
137: PetscInt *size_csr; /* inode sizes in csr with size_csr[0] = 0 and i-th node size = size_csr[i+1] - size_csr[i], to facilitate parallel computation */
138: PetscInt limit; /* inode limit */
139: PetscInt max_limit; /* maximum supported inode limit */
140: PetscBool checked; /* if inodes have been checked for */
141: PetscObjectState mat_nonzerostate; /* non-zero state when inodes were checked for */
142: } Mat_SeqAIJ_Inode;
144: PETSC_INTERN PetscErrorCode MatView_SeqAIJ_Inode(Mat, PetscViewer);
145: PETSC_INTERN PetscErrorCode MatAssemblyEnd_SeqAIJ_Inode(Mat, MatAssemblyType);
146: PETSC_INTERN PetscErrorCode MatDestroy_SeqAIJ_Inode(Mat);
147: PETSC_INTERN PetscErrorCode MatCreate_SeqAIJ_Inode(Mat);
148: PETSC_INTERN PetscErrorCode MatSetOption_SeqAIJ_Inode(Mat, MatOption, PetscBool);
149: PETSC_INTERN PetscErrorCode MatDuplicate_SeqAIJ_Inode(Mat, MatDuplicateOption, Mat *);
150: PETSC_INTERN PetscErrorCode MatDuplicateNoCreate_SeqAIJ(Mat, Mat, MatDuplicateOption, PetscBool);
151: PETSC_INTERN PetscErrorCode MatLUFactorNumeric_SeqAIJ_Inode(Mat, Mat, const MatFactorInfo *);
152: PETSC_INTERN PetscErrorCode MatSeqAIJGetArray_SeqAIJ(Mat, PetscScalar **);
153: PETSC_INTERN PetscErrorCode MatSeqAIJRestoreArray_SeqAIJ(Mat, PetscScalar **);
155: typedef struct {
156: SEQAIJHEADER(MatScalar);
157: Mat_SeqAIJ_Inode inode;
158: MatScalar *saved_values; /* location for stashing nonzero values of matrix */
160: /* data needed for MatSOR_SeqAIJ() */
161: PetscScalar *mdiag, *idiag; /* diagonal values, inverse of diagonal entries */
162: PetscScalar *ssor_work; /* workspace for Eisenstat trick */
163: PetscObjectState idiagState; /* state of the matrix when mdiag and idiag was obtained */
164: PetscScalar fshift, omega; /* last used omega and fshift */
166: PetscScalar *ibdiag; /* inverses of block diagonals */
167: PetscInt ibdiagsize; /* length of ibdiag[], which changes if the block size does */
168: PetscObjectState ibdiagState; /* state of the matrix when ibdiag[] was obtained */
170: /* MatSetValues() via hash related fields */
171: PetscHMapIJV ht;
172: PetscInt *dnz;
173: struct _MatOps cops;
174: } Mat_SeqAIJ;
176: typedef struct {
177: PetscInt nz; /* nz of the matrix after assembly */
178: PetscCount n; /* Number of entries in MatSetPreallocationCOO() */
179: PetscCount Atot; /* Total number of valid (i.e., w/ non-negative indices) entries in the COO array */
180: PetscCount *jmap; /* perm[jmap[i]..jmap[i+1]) give indices of entries in v[] associated with i-th nonzero of the matrix */
181: PetscCount *perm; /* The permutation array in sorting (i,j) by row and then by col */
182: PetscObjectState nonzerostate; /* nonzero state of the matrix that jmap[] indexes into */
183: } MatCOOStruct_SeqAIJ;
185: #define MatSeqXAIJGetOptions_Private(A) \
186: { \
187: const PetscBool oldvalues = (PetscBool)(A != PETSC_NULLPTR); \
188: PetscInt nonew = 0, nounused = 0; \
189: PetscBool roworiented = PETSC_FALSE; \
190: if (oldvalues) { \
191: nonew = ((Mat_SeqAIJ *)(A)->data)->nonew; \
192: nounused = ((Mat_SeqAIJ *)(A)->data)->nounused; \
193: roworiented = ((Mat_SeqAIJ *)(A)->data)->roworiented; \
194: } \
195: (void)0
197: #define MatSeqSBAIJGetOptions_Private(A) \
198: { \
199: PetscBool ignore_ltriangular = PETSC_FALSE, getrow_utriangular = PETSC_FALSE; \
200: MatSeqXAIJGetOptions_Private(A); \
201: if (oldvalues) { \
202: ignore_ltriangular = ((Mat_SeqSBAIJ *)(A)->data)->ignore_ltriangular; \
203: getrow_utriangular = ((Mat_SeqSBAIJ *)(A)->data)->getrow_utriangular; \
204: } \
205: (void)0
207: #define MatSeqXAIJRestoreOptions_Private(A) \
208: if (oldvalues) { \
209: ((Mat_SeqAIJ *)(A)->data)->nonew = nonew; \
210: ((Mat_SeqAIJ *)(A)->data)->nounused = nounused; \
211: ((Mat_SeqAIJ *)(A)->data)->roworiented = roworiented; \
212: } \
213: } \
214: (void)0
216: #define MatSeqSBAIJRestoreOptions_Private(A) \
217: if (oldvalues) { \
218: ((Mat_SeqSBAIJ *)(A)->data)->ignore_ltriangular = ignore_ltriangular; \
219: ((Mat_SeqSBAIJ *)(A)->data)->getrow_utriangular = getrow_utriangular; \
220: } \
221: MatSeqXAIJRestoreOptions_Private(A); \
222: } \
223: (void)0
225: static inline PetscErrorCode MatXAIJAllocatea(Mat A, PetscInt nz, PetscScalar **array)
226: {
227: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
229: PetscFunctionBegin;
230: PetscCall(PetscShmgetAllocateArray(nz, sizeof(PetscScalar), (void **)array));
231: a->free_a = PETSC_TRUE;
232: PetscFunctionReturn(PETSC_SUCCESS);
233: }
235: static inline PetscErrorCode MatXAIJDeallocatea(Mat A, PetscScalar **array)
236: {
237: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
239: PetscFunctionBegin;
240: if (a->free_a) PetscCall(PetscShmgetDeallocateArray((void **)array));
241: a->free_a = PETSC_FALSE;
242: PetscFunctionReturn(PETSC_SUCCESS);
243: }
245: /*
246: Frees the a, i, and j arrays from the XAIJ (AIJ, BAIJ, and SBAIJ) matrix types
247: */
248: static inline PetscErrorCode MatSeqXAIJFreeAIJ(Mat AA, MatScalar **a, PetscInt **j, PetscInt **i)
249: {
250: Mat_SeqAIJ *A = (Mat_SeqAIJ *)AA->data;
252: PetscFunctionBegin;
253: if (A->free_a) PetscCall(PetscShmgetDeallocateArray((void **)a));
254: if (A->free_ij) PetscCall(PetscShmgetDeallocateArray((void **)j));
255: if (A->free_ij) PetscCall(PetscShmgetDeallocateArray((void **)i));
256: PetscFunctionReturn(PETSC_SUCCESS);
257: }
258: /*
259: Allocates larger a, i, and j arrays for the XAIJ (AIJ, BAIJ, and SBAIJ) matrix types
260: This is a macro because it takes the datatype as an argument which can be either a Mat or a MatScalar
261: */
262: #define MatSeqXAIJReallocateAIJ(Amat, AM, BS2, NROW, ROW, COL, RMAX, AA, AI, AJ, RP, AP, AIMAX, NONEW, datatype) \
263: do { \
264: if ((NROW) >= (RMAX)) { \
265: Mat_SeqAIJ *Ain = (Mat_SeqAIJ *)(Amat)->data; \
266: PetscInt CHUNKSIZE = 15, new_nz = (AI)[AM] + CHUNKSIZE, len, *new_i = NULL, *new_j = NULL; \
267: datatype *new_a; \
268: \
269: 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); \
270: /* malloc new storage space */ \
271: PetscCall(PetscShmgetAllocateArray((BS2) * new_nz, sizeof(PetscScalar), (void **)&new_a)); \
272: PetscCall(PetscShmgetAllocateArray(new_nz, sizeof(PetscInt), (void **)&new_j)); \
273: PetscCall(PetscShmgetAllocateArray((AM) + 1, sizeof(PetscInt), (void **)&new_i)); \
274: Ain->free_a = PETSC_TRUE; \
275: Ain->free_ij = PETSC_TRUE; \
276: /* copy over old data into new slots */ \
277: for (ii = 0; ii < (ROW) + 1; ii++) new_i[ii] = (AI)[ii]; \
278: for (ii = (ROW) + 1; ii < (AM) + 1; ii++) new_i[ii] = (AI)[ii] + CHUNKSIZE; \
279: PetscCall(PetscArraycpy(new_j, AJ, (AI)[ROW] + (NROW))); \
280: len = (new_nz - CHUNKSIZE - (AI)[ROW] - (NROW)); \
281: PetscCall(PetscArraycpy(new_j + (AI)[ROW] + (NROW) + CHUNKSIZE, PetscSafePointerPlusOffset(AJ, (AI)[ROW] + (NROW)), len)); \
282: PetscCall(PetscArraycpy(new_a, AA, (BS2) * ((AI)[ROW] + (NROW)))); \
283: PetscCall(PetscArrayzero(new_a + (BS2) * ((AI)[ROW] + (NROW)), (BS2) * CHUNKSIZE)); \
284: PetscCall(PetscArraycpy(new_a + (BS2) * ((AI)[ROW] + (NROW) + CHUNKSIZE), PetscSafePointerPlusOffset(AA, (BS2) * ((AI)[ROW] + (NROW))), (BS2) * len)); \
285: /* free up old matrix storage */ \
286: PetscCall(MatSeqXAIJFreeAIJ(A, &Ain->a, &Ain->j, &Ain->i)); \
287: AA = new_a; \
288: Ain->a = new_a; \
289: AI = Ain->i = new_i; \
290: AJ = Ain->j = new_j; \
291: \
292: RP = (AJ) + (AI)[ROW]; \
293: AP = (AA) + (BS2) * (AI)[ROW]; \
294: RMAX = (AIMAX)[ROW] = (AIMAX)[ROW] + CHUNKSIZE; \
295: Ain->maxnz += (BS2) * CHUNKSIZE; \
296: Ain->reallocs++; \
297: (Amat)->nonzerostate++; \
298: } \
299: } while (0)
301: #define MatSeqXAIJReallocateAIJ_structure_only(Amat, AM, BS2, NROW, ROW, COL, RMAX, AI, AJ, RP, AIMAX, NONEW, datatype) \
302: do { \
303: if ((NROW) >= (RMAX)) { \
304: Mat_SeqAIJ *Ain = (Mat_SeqAIJ *)(Amat)->data; \
305: /* there is no extra room in row, therefore enlarge */ \
306: PetscInt CHUNKSIZE = 15, new_nz = (AI)[AM] + CHUNKSIZE, len, *new_i = NULL, *new_j = NULL; \
307: \
308: 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); \
309: /* malloc new storage space */ \
310: PetscCall(PetscShmgetAllocateArray(new_nz, sizeof(PetscInt), (void **)&new_j)); \
311: PetscCall(PetscShmgetAllocateArray((AM) + 1, sizeof(PetscInt), (void **)&new_i)); \
312: Ain->free_a = PETSC_FALSE; \
313: Ain->free_ij = PETSC_TRUE; \
314: \
315: /* copy over old data into new slots */ \
316: for (ii = 0; ii < (ROW) + 1; ii++) new_i[ii] = (AI)[ii]; \
317: for (ii = (ROW) + 1; ii < (AM) + 1; ii++) new_i[ii] = (AI)[ii] + CHUNKSIZE; \
318: PetscCall(PetscArraycpy(new_j, AJ, (AI)[ROW] + (NROW))); \
319: len = (new_nz - CHUNKSIZE - (AI)[ROW] - (NROW)); \
320: PetscCall(PetscArraycpy(new_j + (AI)[ROW] + (NROW) + CHUNKSIZE, (AJ) + (AI)[ROW] + (NROW), len)); \
321: \
322: /* free up old matrix storage */ \
323: PetscCall(MatSeqXAIJFreeAIJ(A, &Ain->a, &Ain->j, &Ain->i)); \
324: Ain->a = NULL; \
325: AI = Ain->i = new_i; \
326: AJ = Ain->j = new_j; \
327: \
328: RP = (AJ) + (AI)[ROW]; \
329: RMAX = (AIMAX)[ROW] = (AIMAX)[ROW] + CHUNKSIZE; \
330: Ain->maxnz += (BS2) * CHUNKSIZE; \
331: Ain->reallocs++; \
332: (Amat)->nonzerostate++; \
333: } \
334: } while (0)
336: PETSC_INTERN PetscErrorCode MatSeqAIJSetPreallocation_SeqAIJ(Mat, PetscInt, const PetscInt *);
337: PETSC_INTERN PetscErrorCode MatSetPreallocationCOO_SeqAIJ(Mat, PetscCount, PetscInt[], PetscInt[]);
338: PETSC_INTERN PetscErrorCode MatSeqAIJCheckCOONonzeroState(Mat);
340: PETSC_INTERN PetscErrorCode MatILUFactorSymbolic_SeqAIJ(Mat, Mat, IS, IS, const MatFactorInfo *);
341: PETSC_INTERN PetscErrorCode MatILUFactorSymbolic_SeqAIJ_ilu0(Mat, Mat, IS, IS, const MatFactorInfo *);
343: PETSC_INTERN PetscErrorCode MatICCFactorSymbolic_SeqAIJ(Mat, Mat, IS, const MatFactorInfo *);
344: PETSC_INTERN PetscErrorCode MatCholeskyFactorSymbolic_SeqAIJ(Mat, Mat, IS, const MatFactorInfo *);
345: PETSC_INTERN PetscErrorCode MatCholeskyFactorNumeric_SeqAIJ_inplace(Mat, Mat, const MatFactorInfo *);
346: PETSC_INTERN PetscErrorCode MatCholeskyFactorNumeric_SeqAIJ(Mat, Mat, const MatFactorInfo *);
347: PETSC_INTERN PetscErrorCode MatDuplicate_SeqAIJ(Mat, MatDuplicateOption, Mat *);
348: PETSC_INTERN PetscErrorCode MatCopy_SeqAIJ(Mat, Mat, MatStructure);
349: PETSC_EXTERN PetscErrorCode MatGetDiagonalMarkers_SeqAIJ(Mat, const PetscInt **, PetscBool *);
350: PETSC_INTERN PetscErrorCode MatFindZeroDiagonals_SeqAIJ_Private(Mat, PetscInt *, PetscInt **);
352: PETSC_INTERN PetscErrorCode MatMult_SeqAIJ(Mat, Vec, Vec);
353: PETSC_INTERN PetscErrorCode MatMult_SeqAIJ_Inode(Mat, Vec, Vec);
354: PETSC_INTERN PetscErrorCode MatMultAdd_SeqAIJ(Mat, Vec, Vec, Vec);
355: PETSC_INTERN PetscErrorCode MatMultAdd_SeqAIJ_Inode(Mat, Vec, Vec, Vec);
356: PETSC_INTERN PetscErrorCode MatMultTranspose_SeqAIJ(Mat, Vec, Vec);
357: PETSC_INTERN PetscErrorCode MatMultTransposeAdd_SeqAIJ(Mat, Vec, Vec, Vec);
358: PETSC_INTERN PetscErrorCode MatSOR_SeqAIJ(Mat, Vec, PetscReal, MatSORType, PetscReal, PetscInt, PetscInt, Vec);
359: PETSC_INTERN PetscErrorCode MatSOR_SeqAIJ_Inode(Mat, Vec, PetscReal, MatSORType, PetscReal, PetscInt, PetscInt, Vec);
361: PETSC_INTERN PetscErrorCode MatSetOption_SeqAIJ(Mat, MatOption, PetscBool);
363: PETSC_INTERN PetscErrorCode MatGetSymbolicTranspose_SeqAIJ(Mat, PetscInt *[], PetscInt *[]);
364: PETSC_INTERN PetscErrorCode MatRestoreSymbolicTranspose_SeqAIJ(Mat, PetscInt *[], PetscInt *[]);
365: PETSC_INTERN PetscErrorCode MatGetSymbolicTransposeReduced_SeqAIJ(Mat, PetscInt, PetscInt, PetscInt *[], PetscInt *[]);
366: PETSC_INTERN PetscErrorCode MatTransposeSymbolic_SeqAIJ(Mat, Mat *);
367: PETSC_INTERN PetscErrorCode MatTranspose_SeqAIJ(Mat, MatReuse, Mat *);
369: PETSC_INTERN PetscErrorCode MatToSymmetricIJ_SeqAIJ(PetscInt, PetscInt *, PetscInt *, PetscBool, PetscInt, PetscInt, PetscInt **, PetscInt **);
370: PETSC_INTERN PetscErrorCode MatLUFactorSymbolic_SeqAIJ(Mat, Mat, IS, IS, const MatFactorInfo *);
371: PETSC_INTERN PetscErrorCode MatLUFactorNumeric_SeqAIJ_inplace(Mat, Mat, const MatFactorInfo *);
372: PETSC_INTERN PetscErrorCode MatLUFactorNumeric_SeqAIJ(Mat, Mat, const MatFactorInfo *);
373: PETSC_INTERN PetscErrorCode MatLUFactorNumeric_SeqAIJ_InplaceWithPerm(Mat, Mat, const MatFactorInfo *);
374: PETSC_INTERN PetscErrorCode MatLUFactor_SeqAIJ(Mat, IS, IS, const MatFactorInfo *);
375: PETSC_INTERN PetscErrorCode MatSolve_SeqAIJ_inplace(Mat, Vec, Vec);
376: PETSC_INTERN PetscErrorCode MatSolve_SeqAIJ(Mat, Vec, Vec);
377: PETSC_INTERN PetscErrorCode MatSolve_SeqAIJ_Inode(Mat, Vec, Vec);
378: PETSC_INTERN PetscErrorCode MatSolve_SeqAIJ_NaturalOrdering(Mat, Vec, Vec);
379: PETSC_INTERN PetscErrorCode MatSolveAdd_SeqAIJ(Mat, Vec, Vec, Vec);
380: PETSC_INTERN PetscErrorCode MatSolveTranspose_SeqAIJ_inplace(Mat, Vec, Vec);
381: PETSC_INTERN PetscErrorCode MatSolveTranspose_SeqAIJ(Mat, Vec, Vec);
382: PETSC_INTERN PetscErrorCode MatSolveTransposeAdd_SeqAIJ_inplace(Mat, Vec, Vec, Vec);
383: PETSC_INTERN PetscErrorCode MatSolveTransposeAdd_SeqAIJ(Mat, Vec, Vec, Vec);
384: PETSC_INTERN PetscErrorCode MatMatSolve_SeqAIJ(Mat, Mat, Mat);
385: PETSC_INTERN PetscErrorCode MatMatSolveTranspose_SeqAIJ(Mat, Mat, Mat);
386: PETSC_INTERN PetscErrorCode MatEqual_SeqAIJ(Mat, Mat, PetscBool *);
387: PETSC_INTERN PetscErrorCode MatFDColoringCreate_SeqXAIJ(Mat, ISColoring, MatFDColoring);
388: PETSC_INTERN PetscErrorCode MatFDColoringSetUp_SeqXAIJ(Mat, ISColoring, MatFDColoring);
389: PETSC_INTERN PetscErrorCode MatFDColoringSetUpBlocked_AIJ_Private(Mat, MatFDColoring, PetscInt);
390: PETSC_INTERN PetscErrorCode MatLoad_AIJ_HDF5(Mat, PetscViewer);
391: PETSC_INTERN PetscErrorCode MatLoad_SeqAIJ_Binary(Mat, PetscViewer);
392: PETSC_INTERN PetscErrorCode MatLoad_SeqAIJ(Mat, PetscViewer);
394: #if PetscDefined(HAVE_HYPRE)
395: PETSC_INTERN PetscErrorCode MatProductSetFromOptions_Transpose_AIJ_AIJ(Mat);
396: #endif
397: PETSC_INTERN PetscErrorCode MatProductSetFromOptions_SeqAIJ(Mat);
399: PETSC_INTERN PetscErrorCode MatProductSymbolic_PtAP_SeqAIJ_SeqAIJ(Mat);
400: PETSC_INTERN PetscErrorCode MatProductSymbolic_RARt_SeqAIJ_SeqAIJ(Mat);
402: PETSC_INTERN PetscErrorCode MatMatMultSymbolic_SeqAIJ_SeqAIJ(Mat, Mat, PetscReal, Mat);
403: PETSC_INTERN PetscErrorCode MatMatMultSymbolic_SeqAIJ_SeqAIJ_Sorted(Mat, Mat, PetscReal, Mat);
404: PETSC_INTERN PetscErrorCode MatMatMultSymbolic_SeqDense_SeqAIJ(Mat, Mat, PetscReal, Mat);
405: PETSC_INTERN PetscErrorCode MatMatMultSymbolic_SeqAIJ_SeqAIJ_Scalable(Mat, Mat, PetscReal, Mat);
406: PETSC_INTERN PetscErrorCode MatMatMultSymbolic_SeqAIJ_SeqAIJ_Scalable_fast(Mat, Mat, PetscReal, Mat);
407: PETSC_INTERN PetscErrorCode MatMatMultSymbolic_SeqAIJ_SeqAIJ_Heap(Mat, Mat, PetscReal, Mat);
408: PETSC_INTERN PetscErrorCode MatMatMultSymbolic_SeqAIJ_SeqAIJ_BTHeap(Mat, Mat, PetscReal, Mat);
409: PETSC_INTERN PetscErrorCode MatMatMultSymbolic_SeqAIJ_SeqAIJ_RowMerge(Mat, Mat, PetscReal, Mat);
410: PETSC_INTERN PetscErrorCode MatMatMultSymbolic_SeqAIJ_SeqAIJ_LLCondensed(Mat, Mat, PetscReal, Mat);
411: #if PetscDefined(HAVE_HYPRE)
412: PETSC_INTERN PetscErrorCode MatMatMultSymbolic_AIJ_AIJ_wHYPRE(Mat, Mat, PetscReal, Mat);
413: #endif
415: PETSC_INTERN PetscErrorCode MatMatMultNumeric_SeqAIJ_SeqAIJ(Mat, Mat, Mat);
416: PETSC_INTERN PetscErrorCode MatMatMultNumeric_SeqAIJ_SeqAIJ_Sorted(Mat, Mat, Mat);
418: PETSC_INTERN PetscErrorCode MatMatMultNumeric_SeqDense_SeqAIJ(Mat, Mat, Mat);
419: PETSC_INTERN PetscErrorCode MatMatMultNumeric_SeqAIJ_SeqAIJ_Scalable(Mat, Mat, Mat);
421: PETSC_INTERN PetscErrorCode MatPtAPSymbolic_SeqAIJ_SeqAIJ_SparseAxpy(Mat, Mat, PetscReal, Mat);
422: PETSC_INTERN PetscErrorCode MatPtAPNumeric_SeqAIJ_SeqAIJ(Mat, Mat, Mat);
423: PETSC_INTERN PetscErrorCode MatPtAPNumeric_SeqAIJ_SeqAIJ_SparseAxpy(Mat, Mat, Mat);
425: PETSC_INTERN PetscErrorCode MatRARtSymbolic_SeqAIJ_SeqAIJ(Mat, Mat, PetscReal, Mat);
426: PETSC_INTERN PetscErrorCode MatRARtSymbolic_SeqAIJ_SeqAIJ_matmattransposemult(Mat, Mat, PetscReal, Mat);
427: PETSC_INTERN PetscErrorCode MatRARtSymbolic_SeqAIJ_SeqAIJ_colorrart(Mat, Mat, PetscReal, Mat);
428: PETSC_INTERN PetscErrorCode MatRARtNumeric_SeqAIJ_SeqAIJ(Mat, Mat, Mat);
429: PETSC_INTERN PetscErrorCode MatRARtNumeric_SeqAIJ_SeqAIJ_matmattransposemult(Mat, Mat, Mat);
430: PETSC_INTERN PetscErrorCode MatRARtNumeric_SeqAIJ_SeqAIJ_colorrart(Mat, Mat, Mat);
432: PETSC_INTERN PetscErrorCode MatTransposeMatMultSymbolic_SeqAIJ_SeqAIJ(Mat, Mat, PetscReal, Mat);
433: PETSC_INTERN PetscErrorCode MatTransposeMatMultNumeric_SeqAIJ_SeqAIJ(Mat, Mat, Mat);
434: PETSC_INTERN PetscErrorCode MatProductCtxDestroy_SeqAIJ_MatTransMatMult(PetscCtxRt);
436: PETSC_INTERN PetscErrorCode MatMatTransposeMultSymbolic_SeqAIJ_SeqAIJ(Mat, Mat, PetscReal, Mat);
437: PETSC_INTERN PetscErrorCode MatMatTransposeMultNumeric_SeqAIJ_SeqAIJ(Mat, Mat, Mat);
438: PETSC_INTERN PetscErrorCode MatTransposeColoringCreate_SeqAIJ(Mat, ISColoring, MatTransposeColoring);
439: PETSC_INTERN PetscErrorCode MatTransColoringApplySpToDen_SeqAIJ(MatTransposeColoring, Mat, Mat);
440: PETSC_INTERN PetscErrorCode MatTransColoringApplyDenToSp_SeqAIJ(MatTransposeColoring, Mat, Mat);
442: PETSC_INTERN PetscErrorCode MatMatMatMultSymbolic_SeqAIJ_SeqAIJ_SeqAIJ(Mat, Mat, Mat, PetscReal, Mat);
443: PETSC_INTERN PetscErrorCode MatMatMatMultNumeric_SeqAIJ_SeqAIJ_SeqAIJ(Mat, Mat, Mat, Mat);
445: PETSC_INTERN PetscErrorCode MatSetRandomSkipColumnRange_SeqAIJ_Private(Mat, PetscInt, PetscInt, PetscRandom);
446: PETSC_INTERN PetscErrorCode MatSetValues_SeqAIJ(Mat, PetscInt, const PetscInt[], PetscInt, const PetscInt[], const PetscScalar[], InsertMode);
447: PETSC_INTERN PetscErrorCode MatGetRow_SeqAIJ(Mat, PetscInt, PetscInt *, PetscInt **, PetscScalar **);
448: PETSC_INTERN PetscErrorCode MatRestoreRow_SeqAIJ(Mat, PetscInt, PetscInt *, PetscInt **, PetscScalar **);
449: PETSC_INTERN PetscErrorCode MatScale_SeqAIJ(Mat, PetscScalar);
450: PETSC_INTERN PetscErrorCode MatDiagonalScale_SeqAIJ(Mat, Vec, Vec);
451: PETSC_INTERN PetscErrorCode MatDiagonalSet_SeqAIJ(Mat, Vec, InsertMode);
452: PETSC_INTERN PetscErrorCode MatGetRowSumAbs_SeqAIJ(Mat, Vec);
453: PETSC_INTERN PetscErrorCode MatAXPY_SeqAIJ(Mat, PetscScalar, Mat, MatStructure);
454: PETSC_INTERN PetscErrorCode MatGetRowIJ_SeqAIJ(Mat, PetscInt, PetscBool, PetscBool, PetscInt *, const PetscInt *[], const PetscInt *[], PetscBool *);
455: PETSC_INTERN PetscErrorCode MatRestoreRowIJ_SeqAIJ(Mat, PetscInt, PetscBool, PetscBool, PetscInt *, const PetscInt *[], const PetscInt *[], PetscBool *);
456: PETSC_INTERN PetscErrorCode MatGetColumnIJ_SeqAIJ(Mat, PetscInt, PetscBool, PetscBool, PetscInt *, const PetscInt *[], const PetscInt *[], PetscBool *);
457: PETSC_INTERN PetscErrorCode MatRestoreColumnIJ_SeqAIJ(Mat, PetscInt, PetscBool, PetscBool, PetscInt *, const PetscInt *[], const PetscInt *[], PetscBool *);
458: PETSC_INTERN PetscErrorCode MatGetColumnIJ_SeqAIJ_Color(Mat, PetscInt, PetscBool, PetscBool, PetscInt *, const PetscInt *[], const PetscInt *[], PetscInt *[], PetscBool *);
459: PETSC_INTERN PetscErrorCode MatRestoreColumnIJ_SeqAIJ_Color(Mat, PetscInt, PetscBool, PetscBool, PetscInt *, const PetscInt *[], const PetscInt *[], PetscInt *[], PetscBool *);
460: PETSC_INTERN PetscErrorCode MatDestroy_SeqAIJ(Mat);
461: PETSC_INTERN PetscErrorCode MatView_SeqAIJ(Mat, PetscViewer);
463: PETSC_INTERN PetscErrorCode MatSeqAIJCheckInode(Mat);
464: PETSC_INTERN PetscErrorCode MatSeqAIJCheckInode_FactorLU(Mat);
466: PETSC_INTERN PetscErrorCode MatAXPYGetPreallocation_SeqAIJ(Mat, Mat, PetscInt *);
468: #if PetscDefined(HAVE_MATLAB)
469: PETSC_EXTERN PetscErrorCode MatlabEnginePut_SeqAIJ(PetscObject, void *);
470: PETSC_EXTERN PetscErrorCode MatlabEngineGet_SeqAIJ(PetscObject, void *);
471: #endif
472: PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqSBAIJ(Mat, MatType, MatReuse, Mat *);
473: PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqBAIJ(Mat, MatType, MatReuse, Mat *);
474: PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqDense(Mat, MatType, MatReuse, Mat *);
475: PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqAIJCRL(Mat, MatType, MatReuse, Mat *);
476: PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_Elemental(Mat, MatType, MatReuse, Mat *);
477: #if PetscDefined(HAVE_SCALAPACK)
478: PETSC_INTERN PetscErrorCode MatConvert_AIJ_ScaLAPACK(Mat, MatType, MatReuse, Mat *);
479: #endif
480: PETSC_INTERN PetscErrorCode MatConvert_AIJ_HYPRE(Mat, MatType, MatReuse, Mat *);
481: PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqAIJPERM(Mat, MatType, MatReuse, Mat *);
482: PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqAIJSELL(Mat, MatType, MatReuse, Mat *);
483: PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqAIJMKL(Mat, MatType, MatReuse, Mat *);
484: PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqAIJViennaCL(Mat, MatType, MatReuse, Mat *);
485: PETSC_INTERN PetscErrorCode MatReorderForNonzeroDiagonal_SeqAIJ(Mat, PetscReal, IS, IS);
486: PETSC_INTERN PetscErrorCode MatRARt_SeqAIJ_SeqAIJ(Mat, Mat, MatReuse, PetscReal, Mat *);
487: PETSC_EXTERN PetscErrorCode MatCreate_SeqAIJ(Mat);
488: PETSC_INTERN PetscErrorCode MatAssemblyEnd_SeqAIJ(Mat, MatAssemblyType);
489: PETSC_INTERN PetscErrorCode MatZeroEntries_SeqAIJ(Mat);
491: PETSC_INTERN PetscErrorCode MatAXPYGetPreallocation_SeqX_private(PetscInt, const PetscInt *, const PetscInt *, const PetscInt *, const PetscInt *, PetscInt *);
492: PETSC_INTERN PetscErrorCode MatCreateMPIMatConcatenateSeqMat_SeqAIJ(MPI_Comm, Mat, PetscInt, MatReuse, Mat *);
493: PETSC_INTERN PetscErrorCode MatCreateMPIMatConcatenateSeqMat_MPIAIJ(MPI_Comm, Mat, PetscInt, MatReuse, Mat *);
495: PETSC_INTERN PetscErrorCode MatSetSeqMat_SeqAIJ(Mat, IS, IS, MatStructure, Mat);
496: PETSC_INTERN PetscErrorCode MatEliminateZeros_SeqAIJ(Mat, PetscBool);
497: PETSC_INTERN PetscErrorCode MatDestroySubMatrix_Private(Mat_SubSppt *);
498: PETSC_INTERN PetscErrorCode MatDestroySubMatrix_SeqAIJ(Mat);
499: PETSC_INTERN PetscErrorCode MatDestroySubMatrix_Dummy(Mat);
500: PETSC_INTERN PetscErrorCode MatDestroySubMatrices_Dummy(PetscInt, Mat *[]);
501: PETSC_INTERN PetscErrorCode MatCreateSubMatrix_SeqAIJ(Mat, IS, IS, PetscInt, MatReuse, Mat *);
503: PETSC_INTERN PetscErrorCode MatSetSeqAIJWithArrays_private(MPI_Comm, PetscInt, PetscInt, PetscInt[], PetscInt[], PetscScalar[], MatType, Mat);
505: PETSC_INTERN PetscErrorCode MatResetPreallocation_SeqAIJ_Private(Mat A, PetscBool *memoryreset);
507: PETSC_SINGLE_LIBRARY_INTERN PetscErrorCode MatSeqAIJCompactOutExtraColumns_SeqAIJ(Mat, ISLocalToGlobalMapping *);
509: /*
510: PetscSparseDenseMinusDot - The inner kernel of triangular solves and Gauss-Siedel smoothing. \sum_i xv[i] * r[xi[i]] for CSR storage
512: Input Parameters:
513: + nnz - the number of entries
514: . r - the array of vector values
515: . xv - the matrix values for the row
516: - xi - the column indices of the nonzeros in the row
518: Output Parameter:
519: . sum - negative the sum of results
521: PETSc compile flags:
522: + PETSC_KERNEL_USE_UNROLL_4
523: - PETSC_KERNEL_USE_UNROLL_2
525: Developer Note:
526: The macro changes sum but not other parameters
528: .seealso: `PetscSparseDensePlusDot()`
529: */
530: #if PetscDefined(KERNEL_USE_UNROLL_4)
531: #define PetscSparseDenseMinusDot(sum, r, xv, xi, nnz) \
532: do { \
533: if ((nnz) > 0) { \
534: PetscInt nnz2 = nnz, rem = (nnz) & 0x3; \
535: switch (rem) { \
536: case 3: \
537: (sum) -= *(xv)++ * (r)[*(xi)++]; \
538: case 2: \
539: (sum) -= *(xv)++ * (r)[*(xi)++]; \
540: case 1: \
541: (sum) -= *(xv)++ * (r)[*(xi)++]; \
542: nnz2 -= rem; \
543: } \
544: while (nnz2 > 0) { \
545: (sum) -= (xv)[0] * (r)[(xi)[0]] + (xv)[1] * (r)[(xi)[1]] + (xv)[2] * (r)[(xi)[2]] + (xv)[3] * (r)[(xi)[3]]; \
546: (xv) += 4; \
547: (xi) += 4; \
548: nnz2 -= 4; \
549: } \
550: (xv) -= nnz; \
551: (xi) -= nnz; \
552: } \
553: } while (0)
555: #elif PetscDefined(KERNEL_USE_UNROLL_2)
556: #define PetscSparseDenseMinusDot(sum, r, xv, xi, nnz) \
557: do { \
558: PetscInt __i, __i1, __i2; \
559: for (__i = 0; __i < (nnz) - 1; __i += 2) { \
560: __i1 = (xi)[__i]; \
561: __i2 = (xi)[__i + 1]; \
562: (sum) -= ((xv)[__i] * (r)[__i1] + (xv)[__i + 1] * (r)[__i2]); \
563: } \
564: if ((nnz) & 0x1) (sum) -= (xv)[__i] * (r)[(xi)[__i]]; \
565: } while (0)
567: #else
568: #define PetscSparseDenseMinusDot(sum, r, xv, xi, nnz) \
569: do { \
570: PetscInt __i; \
571: for (__i = 0; __i < (nnz); __i++) (sum) -= (xv)[__i] * (r)[(xi)[__i]]; \
572: } while (0)
573: #endif
575: /*
576: PetscSparseDensePlusDot - The inner kernel of matrix-vector product \sum_i xv[i] * r[xi[i]] for CSR storage
578: Input Parameters:
579: + nnz - the number of entries
580: . r - the array of vector values
581: . xv - the matrix values for the row
582: - xi - the column indices of the nonzeros in the row
584: Output Parameter:
585: . sum - the sum of results
587: PETSc compile flags:
588: + PETSC_KERNEL_USE_UNROLL_4
589: - PETSC_KERNEL_USE_UNROLL_2
591: Developer Note:
592: The macro changes sum but not other parameters
594: .seealso: `PetscSparseDenseMinusDot()`
595: */
596: #if PetscDefined(KERNEL_USE_UNROLL_4)
597: #define PetscSparseDensePlusDot(sum, r, xv, xi, nnz) \
598: do { \
599: if ((nnz) > 0) { \
600: PetscInt nnz2 = nnz, rem = (nnz) & 0x3; \
601: switch (rem) { \
602: case 3: \
603: (sum) += *(xv)++ * (r)[*(xi)++]; \
604: case 2: \
605: (sum) += *(xv)++ * (r)[*(xi)++]; \
606: case 1: \
607: (sum) += *(xv)++ * (r)[*(xi)++]; \
608: nnz2 -= rem; \
609: } \
610: while (nnz2 > 0) { \
611: (sum) += (xv)[0] * (r)[(xi)[0]] + (xv)[1] * (r)[(xi)[1]] + (xv)[2] * (r)[(xi)[2]] + (xv)[3] * (r)[(xi)[3]]; \
612: (xv) += 4; \
613: (xi) += 4; \
614: nnz2 -= 4; \
615: } \
616: (xv) -= nnz; \
617: (xi) -= nnz; \
618: } \
619: } while (0)
621: #elif PetscDefined(KERNEL_USE_UNROLL_2)
622: #define PetscSparseDensePlusDot(sum, r, xv, xi, nnz) \
623: do { \
624: PetscInt __i, __i1, __i2; \
625: for (__i = 0; __i < (nnz) - 1; __i += 2) { \
626: __i1 = (xi)[__i]; \
627: __i2 = (xi)[__i + 1]; \
628: (sum) += ((xv)[__i] * (r)[__i1] + (xv)[__i + 1] * (r)[__i2]); \
629: } \
630: if ((nnz) & 0x1) (sum) += (xv)[__i] * (r)[(xi)[__i]]; \
631: } while (0)
633: #elif !(defined(__GNUC__) && defined(_OPENMP)) && PetscDefined(USE_AVX512_KERNELS) && PetscDefined(HAVE_IMMINTRIN_H) && defined(__AVX512F__) && PetscDefined(USE_REAL_DOUBLE) && !PetscDefined(USE_COMPLEX) && !PetscDefined(USE_64BIT_INDICES) && !PetscDefined(SKIP_IMMINTRIN_H_CUDAWORKAROUND)
634: #define PetscSparseDensePlusDot(sum, r, xv, xi, nnz) PetscSparseDensePlusDot_AVX512_Private(&(sum), (r), (xv), (xi), (nnz))
636: #else
637: #define PetscSparseDensePlusDot(sum, r, xv, xi, nnz) \
638: do { \
639: PetscInt __i; \
640: for (__i = 0; __i < (nnz); __i++) (sum) += (xv)[__i] * (r)[(xi)[__i]]; \
641: } while (0)
642: #endif
644: #if PetscDefined(USE_AVX512_KERNELS) && PetscDefined(HAVE_IMMINTRIN_H) && defined(__AVX512F__) && PetscDefined(USE_REAL_DOUBLE) && !PetscDefined(USE_COMPLEX) && !PetscDefined(USE_64BIT_INDICES) && !PetscDefined(SKIP_IMMINTRIN_H_CUDAWORKAROUND)
645: #include <immintrin.h>
646: #if !defined(_MM_SCALE_8)
647: #define _MM_SCALE_8 8
648: #endif
650: static inline void PetscSparseDensePlusDot_AVX512_Private(PetscScalar *sum, const PetscScalar *x, const MatScalar *aa, const PetscInt *aj, PetscInt n)
651: {
652: __m512d vec_x, vec_y, vec_vals;
653: __m256i vec_idx;
654: PetscInt j;
656: vec_y = _mm512_setzero_pd();
657: for (j = 0; j < (n >> 3); j++) {
658: vec_idx = _mm256_loadu_si256((__m256i const *)aj);
659: vec_vals = _mm512_loadu_pd(aa);
660: vec_x = _mm512_i32gather_pd(vec_idx, x, _MM_SCALE_8);
661: vec_y = _mm512_fmadd_pd(vec_x, vec_vals, vec_y);
662: aj += 8;
663: aa += 8;
664: }
665: #if defined(__AVX512VL__)
666: /* masked load requires avx512vl, which is not supported by KNL */
667: if (n & 0x07) {
668: __mmask8 mask;
669: mask = (__mmask8)(0xff >> (8 - (n & 0x07)));
670: vec_idx = _mm256_maskz_loadu_epi32(mask, aj);
671: vec_vals = _mm512_maskz_loadu_pd(mask, aa);
672: vec_x = _mm512_mask_i32gather_pd(_mm512_setzero_pd(), mask, vec_idx, x, _MM_SCALE_8);
673: vec_y = _mm512_mask3_fmadd_pd(vec_x, vec_vals, vec_y, mask);
674: }
675: *sum += _mm512_reduce_add_pd(vec_y);
676: #else
677: *sum += _mm512_reduce_add_pd(vec_y);
678: for (j = 0; j < (n & 0x07); j++) *sum += aa[j] * x[aj[j]];
679: #endif
680: }
681: #endif
683: /*
684: PetscSparseDenseMaxDot - The inner kernel of a modified matrix-vector product \max_i xv[i] * r[xi[i]] for CSR storage
686: Input Parameters:
687: + nnz - the number of entries
688: . r - the array of vector values
689: . xv - the matrix values for the row
690: - xi - the column indices of the nonzeros in the row
692: Output Parameter:
693: . max - the max of results
695: .seealso: `PetscSparseDensePlusDot()`, `PetscSparseDenseMinusDot()`
696: */
697: #define PetscSparseDenseMaxDot(max, r, xv, xi, nnz) \
698: do { \
699: for (PetscInt __i = 0; __i < (nnz); __i++) max = PetscMax(PetscRealPart(max), PetscRealPart((xv)[__i] * (r)[(xi)[__i]])); \
700: } while (0)
702: /*
703: Add column indices into table for counting the max nonzeros of merged rows
704: */
705: #define MatRowMergeMax_SeqAIJ(mat, nrows, ta) \
706: do { \
707: if (mat) { \
708: for (PetscInt _row = 0; _row < (nrows); _row++) { \
709: const PetscInt _nz = (mat)->i[_row + 1] - (mat)->i[_row]; \
710: for (PetscInt _j = 0; _j < _nz; _j++) { \
711: PetscInt *_col = _j + (mat)->j + (mat)->i[_row]; \
712: PetscCall(PetscHMapISet((ta), *_col + 1, 1)); \
713: } \
714: } \
715: } \
716: } while (0)
718: /*
719: Add column indices into table for counting the nonzeros of merged rows
720: */
721: #define MatMergeRows_SeqAIJ(mat, nrows, rows, ta) \
722: do { \
723: for (PetscInt _i = 0; _i < (nrows); _i++) { \
724: const PetscInt _row = (rows)[_i]; \
725: const PetscInt _nz = (mat)->i[_row + 1] - (mat)->i[_row]; \
726: for (PetscInt _j = 0; _j < _nz; _j++) { \
727: PetscInt *_col = _j + (mat)->j + (mat)->i[_row]; \
728: PetscCall(PetscHMapISetWithMode((ta), *_col + 1, 1, INSERT_VALUES)); \
729: } \
730: } \
731: } while (0)