Actual source code: mpiaij.h
petsc-3.7.3 2016-08-01
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;
31: Mat_Merge_SeqsToMPI *merge;
32: PetscErrorCode (*destroy)(Mat);
33: PetscErrorCode (*duplicate)(Mat,MatDuplicateOption,Mat*);
34: } Mat_PtAPMPI;
36: typedef struct {
37: Mat A,B; /* local submatrices: A (diag part),
38: B (off-diag part) */
39: PetscMPIInt size; /* size of communicator */
40: PetscMPIInt rank; /* rank of proc in communicator */
42: /* The following variables are used for matrix assembly */
43: PetscBool donotstash; /* PETSC_TRUE if off processor entries dropped */
44: MPI_Request *send_waits; /* array of send requests */
45: MPI_Request *recv_waits; /* array of receive requests */
46: PetscInt nsends,nrecvs; /* numbers of sends and receives */
47: PetscScalar *svalues,*rvalues; /* sending and receiving data */
48: PetscInt rmax; /* maximum message length */
49: #if defined(PETSC_USE_CTABLE)
50: PetscTable colmap;
51: #else
52: PetscInt *colmap; /* local col number of off-diag col */
53: #endif
54: PetscInt *garray; /* global index of all off-processor columns */
56: /* The following variables are used for matrix-vector products */
57: Vec lvec; /* local vector */
58: Vec diag;
59: VecScatter Mvctx; /* scatter context for vector */
60: PetscBool roworiented; /* if true, row-oriented input, default true */
62: /* The following variables are for MatGetRow() */
63: PetscInt *rowindices; /* column indices for row */
64: PetscScalar *rowvalues; /* nonzero values in row */
65: PetscBool getrowactive; /* indicates MatGetRow(), not restored */
67: /* Used by MatDistribute_MPIAIJ() to allow reuse of previous matrix allocation and nonzero pattern */
68: PetscInt *ld; /* number of entries per row left of diagona block */
70: /* Used by MatMatMult() and MatPtAP() */
71: Mat_PtAPMPI *ptap;
73: /* used by MatMatMatMult() */
74: Mat_MatMatMatMult *matmatmatmult;
76: /* Used by MPICUSP and MPICUSPARSE classes */
77: void * spptr;
79: } Mat_MPIAIJ;
81: PETSC_EXTERN PetscErrorCode MatCreate_MPIAIJ(Mat);
83: PETSC_INTERN PetscErrorCode MatSetColoring_MPIAIJ(Mat,ISColoring);
84: PETSC_INTERN PetscErrorCode MatSetValuesAdifor_MPIAIJ(Mat,PetscInt,void*);
85: PETSC_INTERN PetscErrorCode MatSetUpMultiply_MPIAIJ(Mat);
86: PETSC_INTERN PetscErrorCode MatDisAssemble_MPIAIJ(Mat);
87: PETSC_INTERN PetscErrorCode MatDuplicate_MPIAIJ(Mat,MatDuplicateOption,Mat*);
88: PETSC_INTERN PetscErrorCode MatIncreaseOverlap_MPIAIJ(Mat,PetscInt,IS [],PetscInt);
89: PETSC_INTERN PetscErrorCode MatIncreaseOverlap_MPIAIJ_Scalable(Mat,PetscInt,IS [],PetscInt);
90: PETSC_INTERN PetscErrorCode MatFDColoringCreate_MPIXAIJ(Mat,ISColoring,MatFDColoring);
91: PETSC_INTERN PetscErrorCode MatFDColoringSetUp_MPIXAIJ(Mat,ISColoring,MatFDColoring);
92: PETSC_INTERN PetscErrorCode MatGetSubMatrices_MPIAIJ (Mat,PetscInt,const IS[],const IS[],MatReuse,Mat *[]);
93: PETSC_INTERN PetscErrorCode MatGetSubMatricesMPI_MPIAIJ (Mat,PetscInt,const IS[],const IS[],MatReuse,Mat *[]);
94: PETSC_INTERN PetscErrorCode MatGetSubMatrix_MPIAIJ_All(Mat,MatGetSubMatrixOption,MatReuse,Mat *[]);
97: PETSC_INTERN PetscErrorCode MatGetSubMatrix_MPIAIJ(Mat,IS,IS,MatReuse,Mat*);
98: PETSC_INTERN PetscErrorCode MatGetSubMatrix_MPIAIJ_Private (Mat,IS,IS,PetscInt,MatReuse,Mat*);
99: PETSC_INTERN PetscErrorCode MatGetMultiProcBlock_MPIAIJ(Mat,MPI_Comm,MatReuse,Mat*);
101: PETSC_INTERN PetscErrorCode MatLoad_MPIAIJ(Mat,PetscViewer);
102: PETSC_INTERN PetscErrorCode MatCreateColmap_MPIAIJ_Private(Mat);
103: PETSC_INTERN PetscErrorCode MatMatMult_MPIDense_MPIAIJ(Mat,Mat,MatReuse,PetscReal,Mat*);
104: PETSC_INTERN PetscErrorCode MatMatMult_MPIAIJ_MPIAIJ(Mat,Mat,MatReuse,PetscReal,Mat*);
105: PETSC_INTERN PetscErrorCode MatMatMultSymbolic_MPIAIJ_MPIAIJ_nonscalable(Mat,Mat,PetscReal,Mat*);
106: PETSC_INTERN PetscErrorCode MatMatMultSymbolic_MPIAIJ_MPIAIJ(Mat,Mat,PetscReal,Mat*);
107: PETSC_INTERN PetscErrorCode MatMatMultNumeric_MPIAIJ_MPIAIJ(Mat,Mat,Mat);
108: PETSC_INTERN PetscErrorCode MatMatMultNumeric_MPIAIJ_MPIAIJ_nonscalable(Mat,Mat,Mat);
110: PETSC_INTERN PetscErrorCode MatMatMatMult_MPIAIJ_MPIAIJ_MPIAIJ(Mat,Mat,Mat,MatReuse,PetscReal,Mat*);
111: PETSC_INTERN PetscErrorCode MatMatMatMultSymbolic_MPIAIJ_MPIAIJ_MPIAIJ(Mat,Mat,Mat,PetscReal,Mat*);
112: PETSC_INTERN PetscErrorCode MatMatMatMultNumeric_MPIAIJ_MPIAIJ_MPIAIJ(Mat,Mat,Mat,Mat);
114: PETSC_INTERN PetscErrorCode MatPtAP_MPIAIJ_MPIAIJ(Mat,Mat,MatReuse,PetscReal,Mat*);
115: PETSC_INTERN PetscErrorCode MatPtAPSymbolic_MPIAIJ_MPIAIJ(Mat,Mat,PetscReal,Mat*);
116: PETSC_INTERN PetscErrorCode MatPtAPNumeric_MPIAIJ_MPIAIJ(Mat,Mat,Mat);
118: PETSC_INTERN PetscErrorCode MatPtAPSymbolic_MPIAIJ_MPIAIJ_ptap(Mat,Mat,PetscReal,Mat*);
119: PETSC_INTERN PetscErrorCode MatPtAPNumeric_MPIAIJ_MPIAIJ_ptap(Mat,Mat,Mat);
121: PETSC_INTERN PetscErrorCode MatDestroy_MPIAIJ_PtAP(Mat);
122: PETSC_INTERN PetscErrorCode MatDestroy_MPIAIJ(Mat);
124: PETSC_INTERN PetscErrorCode MatGetBrowsOfAoCols_MPIAIJ(Mat,Mat,MatReuse,PetscInt**,PetscInt**,MatScalar**,Mat*);
125: PETSC_INTERN PetscErrorCode MatSetValues_MPIAIJ(Mat,PetscInt,const PetscInt[],PetscInt,const PetscInt[],const PetscScalar [],InsertMode);
126: PETSC_INTERN PetscErrorCode MatDestroy_MPIAIJ_MatMatMult(Mat);
127: PETSC_INTERN PetscErrorCode PetscContainerDestroy_Mat_MatMatMultMPI(void*);
128: PETSC_INTERN PetscErrorCode MatSetOption_MPIAIJ(Mat,MatOption,PetscBool);
130: PETSC_INTERN PetscErrorCode MatTransposeMatMult_MPIAIJ_MPIAIJ(Mat,Mat,MatReuse,PetscReal,Mat*);
131: PETSC_INTERN PetscErrorCode MatTransposeMatMultSymbolic_MPIAIJ_MPIAIJ_nonscalable(Mat,Mat,PetscReal,Mat*);
132: PETSC_INTERN PetscErrorCode MatTransposeMatMultSymbolic_MPIAIJ_MPIAIJ(Mat,Mat,PetscReal,Mat*);
133: PETSC_INTERN PetscErrorCode MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ(Mat,Mat,Mat);
134: PETSC_INTERN PetscErrorCode MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ_nonscalable(Mat,Mat,Mat);
135: PETSC_INTERN PetscErrorCode MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ_matmatmult(Mat,Mat,Mat);
136: PETSC_INTERN PetscErrorCode MatTransposeMatMult_MPIAIJ_MPIDense(Mat,Mat,MatReuse,PetscReal,Mat*);
137: PETSC_INTERN PetscErrorCode MatTransposeMatMultSymbolic_MPIAIJ_MPIDense(Mat,Mat,PetscReal,Mat*);
138: PETSC_INTERN PetscErrorCode MatTransposeMatMultNumeric_MPIAIJ_MPIDense(Mat,Mat,Mat);
139: PETSC_INTERN PetscErrorCode MatGetSeqNonzeroStructure_MPIAIJ(Mat,Mat*);
141: PETSC_INTERN PetscErrorCode MatSetFromOptions_MPIAIJ(PetscOptionItems*,Mat);
142: PETSC_INTERN PetscErrorCode MatMPIAIJSetPreallocation_MPIAIJ(Mat,PetscInt,const PetscInt[],PetscInt,const PetscInt[]);
144: #if !defined(PETSC_USE_COMPLEX) && !defined(PETSC_USE_REAL_SINGLE) && !defined(PETSC_USE_REAL___FLOAT128)
145: PETSC_INTERN PetscErrorCode MatLUFactorSymbolic_MPIAIJ_TFS(Mat,IS,IS,const MatFactorInfo*,Mat*);
146: #endif
147: PETSC_INTERN PetscErrorCode MatSolve_MPIAIJ(Mat,Vec,Vec);
148: PETSC_INTERN PetscErrorCode MatILUFactor_MPIAIJ(Mat,IS,IS,const MatFactorInfo*);
150: PETSC_INTERN PetscErrorCode MatAXPYGetPreallocation_MPIX_private(PetscInt,const PetscInt*,const PetscInt*,const PetscInt*,const PetscInt*,const PetscInt*,const PetscInt*,PetscInt*);
152: extern PetscErrorCode MatGetDiagonalBlock_MPIAIJ(Mat,Mat*);
153: extern PetscErrorCode MatDiagonalScaleLocal_MPIAIJ(Mat,Vec);
155: PETSC_INTERN PetscErrorCode MatGetSeqMats_MPIAIJ(Mat,Mat*,Mat*);
156: PETSC_INTERN PetscErrorCode MatSetSeqMats_MPIAIJ(Mat,IS,IS,IS,MatStructure,Mat,Mat);
158: /* compute apa = A[i,:]*P = Ad[i,:]*P_loc + Ao*[i,:]*P_oth */
159: #define AProw_nonscalable(i,ad,ao,p_loc,p_oth,apa) \
160: {\
161: PetscInt _anz,_pnz,_j,_k,*_ai,*_aj,_row,*_pi,*_pj; \
162: PetscScalar *_aa,_valtmp,*_pa; \
163: /* diagonal portion of A */\
164: _ai = ad->i;\
165: _anz = _ai[i+1] - _ai[i];\
166: _aj = ad->j + _ai[i];\
167: _aa = ad->a + _ai[i];\
168: for (_j=0; _j<_anz; _j++) {\
169: _row = _aj[_j]; \
170: _pi = p_loc->i; \
171: _pnz = _pi[_row+1] - _pi[_row]; \
172: _pj = p_loc->j + _pi[_row]; \
173: _pa = p_loc->a + _pi[_row]; \
174: /* perform dense axpy */ \
175: _valtmp = _aa[_j]; \
176: for (_k=0; _k<_pnz; _k++) { \
177: apa[_pj[_k]] += _valtmp*_pa[_k]; \
178: } \
179: PetscLogFlops(2.0*_pnz); \
180: } \
181: /* off-diagonal portion of A */ \
182: _ai = ao->i;\
183: _anz = _ai[i+1] - _ai[i]; \
184: _aj = ao->j + _ai[i]; \
185: _aa = ao->a + _ai[i]; \
186: for (_j=0; _j<_anz; _j++) { \
187: _row = _aj[_j]; \
188: _pi = p_oth->i; \
189: _pnz = _pi[_row+1] - _pi[_row]; \
190: _pj = p_oth->j + _pi[_row]; \
191: _pa = p_oth->a + _pi[_row]; \
192: /* perform dense axpy */ \
193: _valtmp = _aa[_j]; \
194: for (_k=0; _k<_pnz; _k++) { \
195: apa[_pj[_k]] += _valtmp*_pa[_k]; \
196: } \
197: PetscLogFlops(2.0*_pnz); \
198: } \
199: }
200: #endif