Actual source code: mpiaij.h


  4: #include <../src/mat/impls/aij/seq/aij.h>

  6: typedef struct { /* used by MatCreateMPIAIJSumSeqAIJ for reusing the merged matrix */
  7:   PetscLayout rowmap;
  8:   PetscInt    **buf_ri,**buf_rj;
  9:   PetscMPIInt *len_s,*len_r,*id_r;    /* array of length of comm->size, store send/recv matrix values */
 10:   PetscMPIInt nsend,nrecv;
 11:   PetscInt    *bi,*bj;    /* i and j array of the local portion of mpi C (matrix product) - rename to ci, cj! */
 12:   PetscInt    *owners_co,*coi,*coj;    /* i and j array of (p->B)^T*A*P - used in the communication */
 13: } Mat_Merge_SeqsToMPI;

 15: typedef struct { /* used by MatPtAPXXX_MPIAIJ_MPIAIJ() and MatMatMultXXX_MPIAIJ_MPIAIJ() */
 16:   PetscInt               *startsj_s,*startsj_r;    /* used by MatGetBrowsOfAoCols_MPIAIJ */
 17:   PetscScalar            *bufa;                    /* used by MatGetBrowsOfAoCols_MPIAIJ */
 18:   Mat                     P_loc,P_oth;             /* partial B_seq -- intend to replace B_seq */
 19:   PetscInt                *api,*apj;               /* symbolic i and j arrays of the local product A_loc*B_seq */
 20:   PetscScalar             *apv;
 21:   MatReuse                reuse;                   /* flag to skip MatGetBrowsOfAoCols_MPIAIJ() and MatMPIAIJGetLocalMat() in 1st call of MatPtAPNumeric_MPIAIJ_MPIAIJ() */
 22:   PetscScalar             *apa;                    /* tmp array for store a row of A*P used in MatMatMult() */
 23:   Mat                     A_loc;                   /* used by MatTransposeMatMult(), contains api and apj */
 24:   ISLocalToGlobalMapping  ltog;                    /* mapping from local column indices to global column indices for A_loc */
 25:   Mat                     Pt;                      /* used by MatTransposeMatMult(), Pt = P^T */
 26:   Mat                     Rd,Ro,AP_loc,C_loc,C_oth;
 27:   PetscInt                algType;                 /* implementation algorithm */
 28:   PetscSF                 sf;                      /* use it to communicate remote part of C */
 29:   PetscInt                *c_othi,*c_rmti;

 31:   Mat_Merge_SeqsToMPI *merge;
 32: } Mat_APMPI;

 34: typedef struct {
 35:   Mat A,B;                             /* local submatrices: A (diag part),
 36:                                            B (off-diag part) */
 37:   PetscMPIInt size;                     /* size of communicator */
 38:   PetscMPIInt rank;                     /* rank of proc in communicator */

 40:   /* The following variables are used for matrix assembly */
 41:   PetscBool   donotstash;               /* PETSC_TRUE if off processor entries dropped */
 42:   MPI_Request *send_waits;              /* array of send requests */
 43:   MPI_Request *recv_waits;              /* array of receive requests */
 44:   PetscInt    nsends,nrecvs;           /* numbers of sends and receives */
 45:   PetscScalar *svalues,*rvalues;       /* sending and receiving data */
 46:   PetscInt    rmax;                     /* maximum message length */
 47: #if defined(PETSC_USE_CTABLE)
 48:   PetscTable colmap;
 49: #else
 50:   PetscInt *colmap;                     /* local col number of off-diag col */
 51: #endif
 52:   PetscInt *garray;                     /* global index of all off-processor columns */

 54:   /* The following variables are used for matrix-vector products */
 55:   Vec        lvec;                 /* local vector */
 56:   Vec        diag;
 57:   VecScatter Mvctx;                /* scatter context for vector */
 58:   PetscBool  roworiented;          /* if true, row-oriented input, default true */

 60:   /* The following variables are for MatGetRow() */
 61:   PetscInt    *rowindices;         /* column indices for row */
 62:   PetscScalar *rowvalues;          /* nonzero values in row */
 63:   PetscBool   getrowactive;        /* indicates MatGetRow(), not restored */

 65:   PetscInt *ld;                    /* number of entries per row left of diagonal block */

 67:   /* Used by device classes */
 68:   void * spptr;

 70: } Mat_MPIAIJ;

 72: PETSC_EXTERN PetscErrorCode MatCreate_MPIAIJ(Mat);

 74: PETSC_INTERN PetscErrorCode MatAssemblyEnd_MPIAIJ(Mat,MatAssemblyType);

 76: PETSC_INTERN PetscErrorCode MatSetUpMultiply_MPIAIJ(Mat);
 77: PETSC_INTERN PetscErrorCode MatDisAssemble_MPIAIJ(Mat);
 78: PETSC_INTERN PetscErrorCode MatDuplicate_MPIAIJ(Mat,MatDuplicateOption,Mat*);
 79: PETSC_INTERN PetscErrorCode MatIncreaseOverlap_MPIAIJ(Mat,PetscInt,IS [],PetscInt);
 80: PETSC_INTERN PetscErrorCode MatIncreaseOverlap_MPIAIJ_Scalable(Mat,PetscInt,IS [],PetscInt);
 81: PETSC_INTERN PetscErrorCode MatFDColoringCreate_MPIXAIJ(Mat,ISColoring,MatFDColoring);
 82: PETSC_INTERN PetscErrorCode MatFDColoringSetUp_MPIXAIJ(Mat,ISColoring,MatFDColoring);
 83: PETSC_INTERN PetscErrorCode MatCreateSubMatrices_MPIAIJ (Mat,PetscInt,const IS[],const IS[],MatReuse,Mat *[]);
 84: PETSC_INTERN PetscErrorCode MatCreateSubMatricesMPI_MPIAIJ (Mat,PetscInt,const IS[],const IS[],MatReuse,Mat *[]);
 85: PETSC_INTERN PetscErrorCode MatCreateSubMatrix_MPIAIJ_All(Mat,MatCreateSubMatrixOption,MatReuse,Mat *[]);
 86: PETSC_INTERN PetscErrorCode MatView_MPIAIJ(Mat,PetscViewer);

 88: PETSC_INTERN PetscErrorCode MatCreateSubMatrix_MPIAIJ(Mat,IS,IS,MatReuse,Mat*);
 89: PETSC_INTERN PetscErrorCode MatCreateSubMatrix_MPIAIJ_nonscalable(Mat,IS,IS,PetscInt,MatReuse,Mat*);
 90: PETSC_INTERN PetscErrorCode MatCreateSubMatrix_MPIAIJ_SameRowDist(Mat,IS,IS,IS,MatReuse,Mat*);
 91: PETSC_INTERN PetscErrorCode MatCreateSubMatrix_MPIAIJ_SameRowColDist(Mat,IS,IS,MatReuse,Mat*);
 92: PETSC_INTERN PetscErrorCode MatGetMultiProcBlock_MPIAIJ(Mat,MPI_Comm,MatReuse,Mat*);

 94: PETSC_INTERN PetscErrorCode MatLoad_MPIAIJ(Mat,PetscViewer);
 95: PETSC_INTERN PetscErrorCode MatLoad_MPIAIJ_Binary(Mat,PetscViewer);
 96: PETSC_INTERN PetscErrorCode MatCreateColmap_MPIAIJ_Private(Mat);

 98: PETSC_INTERN PetscErrorCode MatProductSetFromOptions_MPIAIJ(Mat);
 99: PETSC_INTERN PetscErrorCode MatProductSetFromOptions_MPIAIJBACKEND(Mat);
100: PETSC_INTERN PetscErrorCode MatProductSymbolic_MPIAIJBACKEND(Mat);
101: PETSC_INTERN PetscErrorCode MatProductSymbolic_AB_MPIAIJ_MPIAIJ(Mat);

103: PETSC_INTERN PetscErrorCode MatProductSymbolic_PtAP_MPIAIJ_MPIAIJ(Mat);

105: PETSC_INTERN PetscErrorCode MatProductSymbolic_RARt_MPIAIJ_MPIAIJ(Mat);
106: PETSC_INTERN PetscErrorCode MatProductNumeric_RARt_MPIAIJ_MPIAIJ(Mat);

108: PETSC_INTERN PetscErrorCode MatMatMultSymbolic_MPIAIJ_MPIAIJ_nonscalable(Mat,Mat,PetscReal,Mat);
109: PETSC_INTERN PetscErrorCode MatMatMultSymbolic_MPIAIJ_MPIAIJ_seqMPI(Mat,Mat,PetscReal,Mat);
110: PETSC_INTERN PetscErrorCode MatMatMultSymbolic_MPIAIJ_MPIAIJ(Mat,Mat,PetscReal,Mat);
111: PETSC_INTERN PetscErrorCode MatMatMultNumeric_MPIAIJ_MPIAIJ(Mat,Mat,Mat);
112: PETSC_INTERN PetscErrorCode MatMatMultNumeric_MPIAIJ_MPIAIJ_nonscalable(Mat,Mat,Mat);

114: PETSC_INTERN PetscErrorCode MatMatMatMultSymbolic_MPIAIJ_MPIAIJ_MPIAIJ(Mat,Mat,Mat,PetscReal,Mat);
115: PETSC_INTERN PetscErrorCode MatMatMatMultNumeric_MPIAIJ_MPIAIJ_MPIAIJ(Mat,Mat,Mat,Mat);

117: PETSC_INTERN PetscErrorCode MatPtAPSymbolic_MPIAIJ_MPIAIJ(Mat,Mat,PetscReal,Mat);
118: PETSC_INTERN PetscErrorCode MatPtAPNumeric_MPIAIJ_MPIAIJ(Mat,Mat,Mat);

120: PETSC_INTERN PetscErrorCode MatPtAPSymbolic_MPIAIJ_MPIAIJ_scalable(Mat,Mat,PetscReal,Mat);
121: PETSC_INTERN PetscErrorCode MatPtAPSymbolic_MPIAIJ_MPIAIJ_allatonce(Mat,Mat,PetscReal,Mat);
122: PETSC_INTERN PetscErrorCode MatPtAPSymbolic_MPIAIJ_MPIAIJ_allatonce_merged(Mat,Mat,PetscReal,Mat);
123: PETSC_INTERN PetscErrorCode MatPtAPNumeric_MPIAIJ_MPIAIJ_scalable(Mat,Mat,Mat);
124: PETSC_INTERN PetscErrorCode MatPtAPNumeric_MPIAIJ_MPIAIJ_allatonce(Mat,Mat,Mat);
125: PETSC_INTERN PetscErrorCode MatPtAPNumeric_MPIAIJ_MPIAIJ_allatonce_merged(Mat,Mat,Mat);

127: #if defined(PETSC_HAVE_HYPRE)
128: PETSC_INTERN PetscErrorCode MatPtAPSymbolic_AIJ_AIJ_wHYPRE(Mat,Mat,PetscReal,Mat);
129: #endif
130: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIDense(Mat,MatType,MatReuse,Mat*);
131: #if defined(PETSC_HAVE_SCALAPACK)
132: PETSC_INTERN PetscErrorCode MatConvert_AIJ_ScaLAPACK(Mat,MatType,MatReuse,Mat*);
133: #endif

135: PETSC_INTERN PetscErrorCode MatDestroy_MPIAIJ(Mat);
136: PETSC_INTERN PetscErrorCode MatDestroy_MPIAIJ_PtAP(void*);
137: PETSC_INTERN PetscErrorCode MatDestroy_MPIAIJ_MatMatMult(void*);

139: PETSC_INTERN PetscErrorCode MatGetBrowsOfAoCols_MPIAIJ(Mat,Mat,MatReuse,PetscInt**,PetscInt**,MatScalar**,Mat*);
140: PETSC_INTERN PetscErrorCode MatSetValues_MPIAIJ(Mat,PetscInt,const PetscInt[],PetscInt,const PetscInt[],const PetscScalar [],InsertMode);
141: PETSC_INTERN PetscErrorCode MatSetValues_MPIAIJ_CopyFromCSRFormat(Mat,const PetscInt[],const PetscInt[],const PetscScalar[]);
142: PETSC_INTERN PetscErrorCode MatSetValues_MPIAIJ_CopyFromCSRFormat_Symbolic(Mat,const PetscInt[],const PetscInt[]);
143: PETSC_INTERN PetscErrorCode MatSetOption_MPIAIJ(Mat,MatOption,PetscBool);

145: PETSC_INTERN PetscErrorCode MatTransposeMatMultSymbolic_MPIAIJ_MPIAIJ_nonscalable(Mat,Mat,PetscReal,Mat);
146: PETSC_INTERN PetscErrorCode MatTransposeMatMultSymbolic_MPIAIJ_MPIAIJ(Mat,Mat,PetscReal,Mat);
147: PETSC_INTERN PetscErrorCode MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ(Mat,Mat,Mat);
148: PETSC_INTERN PetscErrorCode MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ_nonscalable(Mat,Mat,Mat);
149: PETSC_INTERN PetscErrorCode MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ_matmatmult(Mat,Mat,Mat);
150: PETSC_INTERN PetscErrorCode MatTransposeMatMultSymbolic_MPIAIJ_MPIDense(Mat,Mat,PetscReal,Mat);
151: PETSC_INTERN PetscErrorCode MatGetSeqNonzeroStructure_MPIAIJ(Mat,Mat*);

153: PETSC_INTERN PetscErrorCode MatSetFromOptions_MPIAIJ(PetscOptionItems*,Mat);
154: PETSC_INTERN PetscErrorCode MatMPIAIJSetPreallocation_MPIAIJ(Mat,PetscInt,const PetscInt[],PetscInt,const PetscInt[]);

156: #if !defined(PETSC_USE_COMPLEX) && !defined(PETSC_USE_REAL_SINGLE) && !defined(PETSC_USE_REAL___FLOAT128) && !defined(PETSC_USE_REAL___FP16)
157: PETSC_INTERN PetscErrorCode MatLUFactorSymbolic_MPIAIJ_TFS(Mat,IS,IS,const MatFactorInfo*,Mat*);
158: #endif
159: PETSC_INTERN PetscErrorCode MatSolve_MPIAIJ(Mat,Vec,Vec);
160: PETSC_INTERN PetscErrorCode MatILUFactor_MPIAIJ(Mat,IS,IS,const MatFactorInfo*);

162: PETSC_INTERN PetscErrorCode MatAXPYGetPreallocation_MPIX_private(PetscInt,const PetscInt*,const PetscInt*,const PetscInt*,const PetscInt*,const PetscInt*,const PetscInt*,PetscInt*);

164: extern PetscErrorCode MatGetDiagonalBlock_MPIAIJ(Mat,Mat*);
165: extern PetscErrorCode MatDiagonalScaleLocal_MPIAIJ(Mat,Vec);

167: PETSC_INTERN PetscErrorCode MatGetSeqMats_MPIAIJ(Mat,Mat*,Mat*);
168: PETSC_INTERN PetscErrorCode MatSetSeqMats_MPIAIJ(Mat,IS,IS,IS,MatStructure,Mat,Mat);

170: /* compute apa = A[i,:]*P = Ad[i,:]*P_loc + Ao*[i,:]*P_oth using sparse axpy */
171: #define AProw_scalable(i,ad,ao,p_loc,p_oth,api,apj,apa) \
172: {\
173:   PetscInt    _anz,_pnz,_j,_k,*_ai,*_aj,_row,*_pi,*_pj,_nextp,*_apJ;\
174:   PetscScalar *_aa,_valtmp,*_pa;\
175:   _apJ = apj + api[i];\
176:   /* diagonal portion of A */\
177:   _ai  = ad->i;\
178:   _anz = _ai[i+1] - _ai[i];\
179:   _aj  = ad->j + _ai[i];\
180:   _aa  = ad->a + _ai[i];\
181:   for (_j=0; _j<_anz; _j++) {\
182:     _row = _aj[_j]; \
183:     _pi  = p_loc->i;                             \
184:     _pnz = _pi[_row+1] - _pi[_row];              \
185:     _pj  = p_loc->j + _pi[_row];                 \
186:     _pa  = p_loc->a + _pi[_row];                 \
187:     /* perform sparse axpy */                    \
188:     _valtmp = _aa[_j];                           \
189:     _nextp  = 0; \
190:     for (_k=0; _nextp<_pnz; _k++) {                    \
191:       if (_apJ[_k] == _pj[_nextp]) { /* column of AP == column of P */\
192:         apa[_k] += _valtmp*_pa[_nextp++];                             \
193:       } \
194:     }                                           \
195:     (void)PetscLogFlops(2.0*_pnz);              \
196:   }                                             \
197:   /* off-diagonal portion of A */               \
198:   if (p_oth){ \
199:     _ai  = ao->i;\
200:     _anz = _ai[i+1] - _ai[i];                   \
201:     _aj  = ao->j + _ai[i];                      \
202:     _aa  = ao->a + _ai[i];                      \
203:     for (_j=0; _j<_anz; _j++) {                 \
204:       _row = _aj[_j];    \
205:       _pi  = p_oth->i;                         \
206:       _pnz = _pi[_row+1] - _pi[_row];          \
207:       _pj  = p_oth->j + _pi[_row];             \
208:       _pa  = p_oth->a + _pi[_row];             \
209:       /* perform sparse axpy */                \
210:       _valtmp = _aa[_j];                       \
211:       _nextp  = 0; \
212:       for (_k=0; _nextp<_pnz; _k++) {          \
213:         if (_apJ[_k] == _pj[_nextp]) { /* column of AP == column of P */\
214:           apa[_k] += _valtmp*_pa[_nextp++];    \
215:         }                                      \
216:       }                                        \
217:       (void)PetscLogFlops(2.0*_pnz);           \
218:     } \
219:   }\
220: }

222: #define AProw_nonscalable(i,ad,ao,p_loc,p_oth,apa) \
223: {\
224:   PetscInt    _anz,_pnz,_j,_k,*_ai,*_aj,_row,*_pi,*_pj;\
225:   PetscScalar *_aa,_valtmp,*_pa;                       \
226:   /* diagonal portion of A */\
227:   _ai  = ad->i;\
228:   _anz = _ai[i+1] - _ai[i];\
229:   _aj  = ad->j + _ai[i];\
230:   _aa  = ad->a + _ai[i];\
231:   for (_j=0; _j<_anz; _j++) {\
232:     _row = _aj[_j]; \
233:     _pi  = p_loc->i;                        \
234:     _pnz = _pi[_row+1] - _pi[_row];         \
235:     _pj  = p_loc->j + _pi[_row];            \
236:     _pa  = p_loc->a + _pi[_row];            \
237:     /* perform dense axpy */                \
238:     _valtmp = _aa[_j];                      \
239:     for (_k=0; _k<_pnz; _k++) {             \
240:       apa[_pj[_k]] += _valtmp*_pa[_k];      \
241:     }                                       \
242:     (void)PetscLogFlops(2.0*_pnz);          \
243:   }                                         \
244:   /* off-diagonal portion of A */           \
245:   if (p_oth){ \
246:     _ai  = ao->i;\
247:     _anz = _ai[i+1] - _ai[i];               \
248:     _aj  = ao->j + _ai[i];                  \
249:     _aa  = ao->a + _ai[i];                  \
250:     for (_j=0; _j<_anz; _j++) {             \
251:       _row = _aj[_j];    \
252:       _pi  = p_oth->i;                      \
253:       _pnz = _pi[_row+1] - _pi[_row];       \
254:       _pj  = p_oth->j + _pi[_row];          \
255:       _pa  = p_oth->a + _pi[_row];          \
256:       /* perform dense axpy */              \
257:       _valtmp = _aa[_j];                    \
258:       for (_k=0; _k<_pnz; _k++) {           \
259:         apa[_pj[_k]] += _valtmp*_pa[_k];    \
260:       }                                     \
261:       (void)PetscLogFlops(2.0*_pnz);        \
262:     }                                       \
263:   }\
264: }

266: #endif