Actual source code: mpiaij.h
petsc-3.9.4 2018-09-11
5: #include <../src/mat/impls/aij/seq/aij.h>
7: typedef struct { /* used by MatCreateMPIAIJSumSeqAIJ for reusing the merged matrix */
8: PetscLayout rowmap;
9: PetscInt **buf_ri,**buf_rj;
10: PetscMPIInt *len_s,*len_r,*id_r; /* array of length of comm->size, store send/recv matrix values */
11: PetscMPIInt nsend,nrecv;
12: PetscInt *bi,*bj; /* i and j array of the local portion of mpi C (matrix product) - rename to ci, cj! */
13: PetscInt *owners_co,*coi,*coj; /* i and j array of (p->B)^T*A*P - used in the communication */
14: PetscErrorCode (*destroy)(Mat);
15: PetscErrorCode (*duplicate)(Mat,MatDuplicateOption,Mat*);
16: } Mat_Merge_SeqsToMPI;
18: typedef struct { /* used by MatPtAP_MPIAIJ_MPIAIJ() and MatMatMult_MPIAIJ_MPIAIJ() */
19: PetscInt *startsj_s,*startsj_r; /* used by MatGetBrowsOfAoCols_MPIAIJ */
20: PetscScalar *bufa; /* used by MatGetBrowsOfAoCols_MPIAIJ */
21: Mat P_loc,P_oth; /* partial B_seq -- intend to replace B_seq */
22: PetscInt *api,*apj; /* symbolic i and j arrays of the local product A_loc*B_seq */
23: PetscScalar *apv;
24: MatReuse reuse; /* flag to skip MatGetBrowsOfAoCols_MPIAIJ() and MatMPIAIJGetLocalMat() in 1st call of MatPtAPNumeric_MPIAIJ_MPIAIJ() */
25: PetscScalar *apa; /* tmp array for store a row of A*P used in MatMatMult() */
26: Mat A_loc; /* used by MatTransposeMatMult(), contains api and apj */
27: Mat Pt; /* used by MatTransposeMatMult(), Pt = P^T */
28: PetscBool scalable; /* flag determines scalable or non-scalable implementation */
29: Mat Rd,Ro,AP_loc,C_loc,C_oth;
30: PetscInt algType; /* implementation algorithm */
32: Mat_Merge_SeqsToMPI *merge;
33: PetscErrorCode (*destroy)(Mat);
34: PetscErrorCode (*duplicate)(Mat,MatDuplicateOption,Mat*);
35: PetscErrorCode (*view)(Mat,PetscViewer);
36: } Mat_PtAPMPI;
38: typedef struct {
39: Mat A,B; /* local submatrices: A (diag part),
40: B (off-diag part) */
41: PetscMPIInt size; /* size of communicator */
42: PetscMPIInt rank; /* rank of proc in communicator */
44: /* The following variables are used for matrix assembly */
45: PetscBool donotstash; /* PETSC_TRUE if off processor entries dropped */
46: MPI_Request *send_waits; /* array of send requests */
47: MPI_Request *recv_waits; /* array of receive requests */
48: PetscInt nsends,nrecvs; /* numbers of sends and receives */
49: PetscScalar *svalues,*rvalues; /* sending and receiving data */
50: PetscInt rmax; /* maximum message length */
51: #if defined(PETSC_USE_CTABLE)
52: PetscTable colmap;
53: #else
54: PetscInt *colmap; /* local col number of off-diag col */
55: #endif
56: PetscInt *garray; /* global index of all off-processor columns */
58: /* The following variables are used for matrix-vector products */
59: Vec lvec; /* local vector */
60: Vec diag;
61: VecScatter Mvctx,Mvctx_mpi1; /* scatter context for vector */
62: PetscBool Mvctx_mpi1_flg; /* if true, additional Mvctx_mpi1 is requested for mat-mat ops, default false */
63: PetscBool roworiented; /* if true, row-oriented input, default true */
65: /* The following variables are for MatGetRow() */
66: PetscInt *rowindices; /* column indices for row */
67: PetscScalar *rowvalues; /* nonzero values in row */
68: PetscBool getrowactive; /* indicates MatGetRow(), not restored */
70: /* Used by MatDistribute_MPIAIJ() to allow reuse of previous matrix allocation and nonzero pattern */
71: PetscInt *ld; /* number of entries per row left of diagona block */
73: /* Used by MatMatMult() and MatPtAP() */
74: Mat_PtAPMPI *ptap;
76: /* used by MatMatMatMult() */
77: Mat_MatMatMatMult *matmatmatmult;
79: /* Used by MPICUSP and MPICUSPARSE classes */
80: void * spptr;
82: } Mat_MPIAIJ;
84: PETSC_EXTERN PetscErrorCode MatCreate_MPIAIJ(Mat);
86: PETSC_INTERN PetscErrorCode MatAssemblyEnd_MPIAIJ(Mat,MatAssemblyType);
88: PETSC_INTERN PetscErrorCode MatSetUpMultiply_MPIAIJ(Mat);
89: PETSC_INTERN PetscErrorCode MatDisAssemble_MPIAIJ(Mat);
90: PETSC_INTERN PetscErrorCode MatDuplicate_MPIAIJ(Mat,MatDuplicateOption,Mat*);
91: PETSC_INTERN PetscErrorCode MatIncreaseOverlap_MPIAIJ(Mat,PetscInt,IS [],PetscInt);
92: PETSC_INTERN PetscErrorCode MatIncreaseOverlap_MPIAIJ_Scalable(Mat,PetscInt,IS [],PetscInt);
93: PETSC_INTERN PetscErrorCode MatFDColoringCreate_MPIXAIJ(Mat,ISColoring,MatFDColoring);
94: PETSC_INTERN PetscErrorCode MatFDColoringSetUp_MPIXAIJ(Mat,ISColoring,MatFDColoring);
95: PETSC_INTERN PetscErrorCode MatCreateSubMatrices_MPIAIJ (Mat,PetscInt,const IS[],const IS[],MatReuse,Mat *[]);
96: PETSC_INTERN PetscErrorCode MatCreateSubMatricesMPI_MPIAIJ (Mat,PetscInt,const IS[],const IS[],MatReuse,Mat *[]);
97: PETSC_INTERN PetscErrorCode MatCreateSubMatrix_MPIAIJ_All(Mat,MatCreateSubMatrixOption,MatReuse,Mat *[]);
100: PETSC_INTERN PetscErrorCode MatCreateSubMatrix_MPIAIJ(Mat,IS,IS,MatReuse,Mat*);
101: PETSC_INTERN PetscErrorCode MatCreateSubMatrix_MPIAIJ_nonscalable(Mat,IS,IS,PetscInt,MatReuse,Mat*);
102: PETSC_INTERN PetscErrorCode MatCreateSubMatrix_MPIAIJ_SameRowDist(Mat,IS,IS,IS,MatReuse,Mat*);
103: PETSC_INTERN PetscErrorCode MatCreateSubMatrix_MPIAIJ_SameRowColDist(Mat,IS,IS,MatReuse,Mat*);
104: PETSC_INTERN PetscErrorCode MatGetMultiProcBlock_MPIAIJ(Mat,MPI_Comm,MatReuse,Mat*);
106: PETSC_INTERN PetscErrorCode MatLoad_MPIAIJ(Mat,PetscViewer);
107: PETSC_INTERN PetscErrorCode MatCreateColmap_MPIAIJ_Private(Mat);
108: PETSC_INTERN PetscErrorCode MatMatMult_MPIDense_MPIAIJ(Mat,Mat,MatReuse,PetscReal,Mat*);
109: PETSC_INTERN PetscErrorCode MatMatMult_MPIAIJ_MPIAIJ(Mat,Mat,MatReuse,PetscReal,Mat*);
110: PETSC_INTERN PetscErrorCode MatMatMultSymbolic_MPIAIJ_MPIAIJ_nonscalable(Mat,Mat,PetscReal,Mat*);
111: PETSC_INTERN PetscErrorCode MatMatMultSymbolic_MPIAIJ_MPIAIJ(Mat,Mat,PetscReal,Mat*);
112: PETSC_INTERN PetscErrorCode MatMatMultNumeric_MPIAIJ_MPIAIJ(Mat,Mat,Mat);
113: PETSC_INTERN PetscErrorCode MatMatMultNumeric_MPIAIJ_MPIAIJ_nonscalable(Mat,Mat,Mat);
115: PETSC_INTERN PetscErrorCode MatMatMatMult_MPIAIJ_MPIAIJ_MPIAIJ(Mat,Mat,Mat,MatReuse,PetscReal,Mat*);
116: PETSC_INTERN PetscErrorCode MatMatMatMultSymbolic_MPIAIJ_MPIAIJ_MPIAIJ(Mat,Mat,Mat,PetscReal,Mat*);
117: PETSC_INTERN PetscErrorCode MatMatMatMultNumeric_MPIAIJ_MPIAIJ_MPIAIJ(Mat,Mat,Mat,Mat);
119: PETSC_INTERN PetscErrorCode MatPtAP_MPIAIJ_MPIAIJ(Mat,Mat,MatReuse,PetscReal,Mat*);
120: PETSC_INTERN PetscErrorCode MatPtAPSymbolic_MPIAIJ_MPIAIJ(Mat,Mat,PetscReal,Mat*);
121: PETSC_INTERN PetscErrorCode MatPtAPNumeric_MPIAIJ_MPIAIJ(Mat,Mat,Mat);
123: PETSC_INTERN PetscErrorCode MatPtAPSymbolic_MPIAIJ_MPIAIJ_scalable(Mat,Mat,PetscReal,Mat*);
124: PETSC_INTERN PetscErrorCode MatPtAPNumeric_MPIAIJ_MPIAIJ_scalable(Mat,Mat,Mat);
125: #if defined(PETSC_HAVE_HYPRE)
126: PETSC_INTERN PetscErrorCode MatPtAPSymbolic_AIJ_AIJ_wHYPRE(Mat,Mat,PetscReal,Mat*);
127: #endif
129: PETSC_INTERN PetscErrorCode MatDestroy_MPIAIJ_PtAP(Mat);
130: PETSC_INTERN PetscErrorCode MatDestroy_MPIAIJ(Mat);
132: PETSC_INTERN PetscErrorCode MatRARt_MPIAIJ_MPIAIJ(Mat,Mat,MatReuse,PetscReal,Mat*);
134: PETSC_INTERN PetscErrorCode MatGetBrowsOfAoCols_MPIAIJ(Mat,Mat,MatReuse,PetscInt**,PetscInt**,MatScalar**,Mat*);
135: PETSC_INTERN PetscErrorCode MatSetValues_MPIAIJ(Mat,PetscInt,const PetscInt[],PetscInt,const PetscInt[],const PetscScalar [],InsertMode);
136: PETSC_INTERN PetscErrorCode MatDestroy_MPIAIJ_MatMatMult(Mat);
137: PETSC_INTERN PetscErrorCode PetscContainerDestroy_Mat_MatMatMultMPI(void*);
138: PETSC_INTERN PetscErrorCode MatSetOption_MPIAIJ(Mat,MatOption,PetscBool);
140: PETSC_INTERN PetscErrorCode MatTransposeMatMult_MPIAIJ_MPIAIJ(Mat,Mat,MatReuse,PetscReal,Mat*);
141: PETSC_INTERN PetscErrorCode MatTransposeMatMultSymbolic_MPIAIJ_MPIAIJ_nonscalable(Mat,Mat,PetscReal,Mat*);
142: PETSC_INTERN PetscErrorCode MatTransposeMatMultSymbolic_MPIAIJ_MPIAIJ(Mat,Mat,PetscReal,Mat*);
143: PETSC_INTERN PetscErrorCode MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ(Mat,Mat,Mat);
144: PETSC_INTERN PetscErrorCode MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ_nonscalable(Mat,Mat,Mat);
145: PETSC_INTERN PetscErrorCode MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ_matmatmult(Mat,Mat,Mat);
146: PETSC_INTERN PetscErrorCode MatTransposeMatMult_MPIAIJ_MPIDense(Mat,Mat,MatReuse,PetscReal,Mat*);
147: PETSC_INTERN PetscErrorCode MatTransposeMatMultSymbolic_MPIAIJ_MPIDense(Mat,Mat,PetscReal,Mat*);
148: PETSC_INTERN PetscErrorCode MatTransposeMatMultNumeric_MPIAIJ_MPIDense(Mat,Mat,Mat);
149: PETSC_INTERN PetscErrorCode MatGetSeqNonzeroStructure_MPIAIJ(Mat,Mat*);
151: PETSC_INTERN PetscErrorCode MatSetFromOptions_MPIAIJ(PetscOptionItems*,Mat);
152: PETSC_INTERN PetscErrorCode MatMPIAIJSetPreallocation_MPIAIJ(Mat,PetscInt,const PetscInt[],PetscInt,const PetscInt[]);
154: #if !defined(PETSC_USE_COMPLEX) && !defined(PETSC_USE_REAL_SINGLE) && !defined(PETSC_USE_REAL___FLOAT128) && !defined(PETSC_USE_REAL___FP16)
155: PETSC_INTERN PetscErrorCode MatLUFactorSymbolic_MPIAIJ_TFS(Mat,IS,IS,const MatFactorInfo*,Mat*);
156: #endif
157: PETSC_INTERN PetscErrorCode MatSolve_MPIAIJ(Mat,Vec,Vec);
158: PETSC_INTERN PetscErrorCode MatILUFactor_MPIAIJ(Mat,IS,IS,const MatFactorInfo*);
160: PETSC_INTERN PetscErrorCode MatAXPYGetPreallocation_MPIX_private(PetscInt,const PetscInt*,const PetscInt*,const PetscInt*,const PetscInt*,const PetscInt*,const PetscInt*,PetscInt*);
162: extern PetscErrorCode MatGetDiagonalBlock_MPIAIJ(Mat,Mat*);
163: extern PetscErrorCode MatDiagonalScaleLocal_MPIAIJ(Mat,Vec);
165: PETSC_INTERN PetscErrorCode MatGetSeqMats_MPIAIJ(Mat,Mat*,Mat*);
166: PETSC_INTERN PetscErrorCode MatSetSeqMats_MPIAIJ(Mat,IS,IS,IS,MatStructure,Mat,Mat);
168: /* compute apa = A[i,:]*P = Ad[i,:]*P_loc + Ao*[i,:]*P_oth using sparse axpy */
169: #define AProw_scalable(i,ad,ao,p_loc,p_oth,api,apj,apa) \
170: {\
171: PetscInt _anz,_pnz,_j,_k,*_ai,*_aj,_row,*_pi,*_pj,_nextp,*_apJ; \
172: PetscScalar *_aa,_valtmp,*_pa; \
173: _apJ = apj + api[i];\
174: /* diagonal portion of A */\
175: _ai = ad->i;\
176: _anz = _ai[i+1] - _ai[i];\
177: _aj = ad->j + _ai[i];\
178: _aa = ad->a + _ai[i];\
179: for (_j=0; _j<_anz; _j++) {\
180: _row = _aj[_j]; \
181: _pi = p_loc->i; \
182: _pnz = _pi[_row+1] - _pi[_row]; \
183: _pj = p_loc->j + _pi[_row]; \
184: _pa = p_loc->a + _pi[_row]; \
185: /* perform sparse axpy */ \
186: _valtmp = _aa[_j]; \
187: _nextp = 0; \
188: for (_k=0; _nextp<_pnz; _k++) { \
189: if (_apJ[_k] == _pj[_nextp]) { /* column of AP == column of P */ \
190: apa[_k] += _valtmp*_pa[_nextp++]; \
191: } \
192: } \
193: (void)PetscLogFlops(2.0*_pnz); \
194: } \
195: /* off-diagonal portion of A */ \
196: _ai = ao->i;\
197: _anz = _ai[i+1] - _ai[i]; \
198: _aj = ao->j + _ai[i]; \
199: _aa = ao->a + _ai[i]; \
200: for (_j=0; _j<_anz; _j++) { \
201: _row = _aj[_j]; \
202: _pi = p_oth->i; \
203: _pnz = _pi[_row+1] - _pi[_row]; \
204: _pj = p_oth->j + _pi[_row]; \
205: _pa = p_oth->a + _pi[_row]; \
206: /* perform sparse axpy */ \
207: _valtmp = _aa[_j]; \
208: _nextp = 0; \
209: for (_k=0; _nextp<_pnz; _k++) { \
210: if (_apJ[_k] == _pj[_nextp]) { /* column of AP == column of P */\
211: apa[_k] += _valtmp*_pa[_nextp++]; \
212: } \
213: } \
214: (void)PetscLogFlops(2.0*_pnz); \
215: } \
216: }
218: #define AProw_nonscalable(i,ad,ao,p_loc,p_oth,apa) \
219: {\
220: PetscInt _anz,_pnz,_j,_k,*_ai,*_aj,_row,*_pi,*_pj; \
221: PetscScalar *_aa,_valtmp,*_pa; \
222: /* diagonal portion of A */\
223: _ai = ad->i;\
224: _anz = _ai[i+1] - _ai[i];\
225: _aj = ad->j + _ai[i];\
226: _aa = ad->a + _ai[i];\
227: for (_j=0; _j<_anz; _j++) {\
228: _row = _aj[_j]; \
229: _pi = p_loc->i; \
230: _pnz = _pi[_row+1] - _pi[_row]; \
231: _pj = p_loc->j + _pi[_row]; \
232: _pa = p_loc->a + _pi[_row]; \
233: /* perform dense axpy */ \
234: _valtmp = _aa[_j]; \
235: for (_k=0; _k<_pnz; _k++) { \
236: apa[_pj[_k]] += _valtmp*_pa[_k]; \
237: } \
238: (void)PetscLogFlops(2.0*_pnz); \
239: } \
240: /* off-diagonal portion of A */ \
241: _ai = ao->i;\
242: _anz = _ai[i+1] - _ai[i]; \
243: _aj = ao->j + _ai[i]; \
244: _aa = ao->a + _ai[i]; \
245: for (_j=0; _j<_anz; _j++) { \
246: _row = _aj[_j]; \
247: _pi = p_oth->i; \
248: _pnz = _pi[_row+1] - _pi[_row]; \
249: _pj = p_oth->j + _pi[_row]; \
250: _pa = p_oth->a + _pi[_row]; \
251: /* perform dense axpy */ \
252: _valtmp = _aa[_j]; \
253: for (_k=0; _k<_pnz; _k++) { \
254: apa[_pj[_k]] += _valtmp*_pa[_k]; \
255: } \
256: (void)PetscLogFlops(2.0*_pnz); \
257: } \
258: }
260: #endif