1 2 /* 3 Defines matrix-matrix product routines for pairs of MPIAIJ matrices 4 C = A * B 5 */ 6 #include <../src/mat/impls/aij/seq/aij.h> /*I "petscmat.h" I*/ 7 #include <../src/mat/utils/freespace.h> 8 #include <../src/mat/impls/aij/mpi/mpiaij.h> 9 #include <petscbt.h> 10 #include <../src/mat/impls/dense/mpi/mpidense.h> 11 /* 12 #define DEBUG_MATMATMULT 13 */ 14 /* 15 #define DEBUG_MATTrMatMult 16 */ 17 18 #undef __FUNCT__ 19 #define __FUNCT__ "MatMatMult_MPIAIJ_MPIAIJ" 20 PetscErrorCode MatMatMult_MPIAIJ_MPIAIJ(Mat A,Mat B,MatReuse scall,PetscReal fill, Mat *C) 21 { 22 PetscErrorCode ierr; 23 24 PetscFunctionBegin; 25 if (scall == MAT_INITIAL_MATRIX){ 26 ierr = PetscLogEventBegin(MAT_MatMultSymbolic,A,B,0,0);CHKERRQ(ierr); 27 ierr = MatMatMultSymbolic_MPIAIJ_MPIAIJ(A,B,fill,C);CHKERRQ(ierr); 28 ierr = PetscLogEventEnd(MAT_MatMultSymbolic,A,B,0,0);CHKERRQ(ierr); 29 } 30 ierr = PetscLogEventBegin(MAT_MatMultNumeric,A,B,0,0);CHKERRQ(ierr); 31 ierr = (*(*C)->ops->matmultnumeric)(A,B,*C);CHKERRQ(ierr); 32 ierr = PetscLogEventEnd(MAT_MatMultNumeric,A,B,0,0);CHKERRQ(ierr); 33 PetscFunctionReturn(0); 34 } 35 36 #undef __FUNCT__ 37 #define __FUNCT__ "PetscContainerDestroy_Mat_MatMatMultMPI" 38 PetscErrorCode PetscContainerDestroy_Mat_MatMatMultMPI(void *ptr) 39 { 40 PetscErrorCode ierr; 41 Mat_MatMatMultMPI *mult=(Mat_MatMatMultMPI*)ptr; 42 43 PetscFunctionBegin; 44 ierr = ISDestroy(&mult->isrowa);CHKERRQ(ierr); 45 ierr = ISDestroy(&mult->isrowb);CHKERRQ(ierr); 46 ierr = ISDestroy(&mult->iscolb);CHKERRQ(ierr); 47 ierr = MatDestroy(&mult->C_seq);CHKERRQ(ierr); 48 ierr = MatDestroy(&mult->A_loc);CHKERRQ(ierr); 49 ierr = MatDestroy(&mult->B_seq);CHKERRQ(ierr); 50 ierr = PetscFree(mult);CHKERRQ(ierr); 51 PetscFunctionReturn(0); 52 } 53 54 #undef __FUNCT__ 55 #define __FUNCT__ "MatDestroy_MPIAIJ_MatMatMult" 56 PetscErrorCode MatDestroy_MPIAIJ_MatMatMult(Mat A) 57 { 58 PetscErrorCode ierr; 59 Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data; 60 Mat_PtAPMPI *ptap=a->ptap; 61 62 PetscFunctionBegin; 63 ierr = PetscFree2(ptap->startsj,ptap->startsj_r);CHKERRQ(ierr); 64 ierr = PetscFree(ptap->bufa);CHKERRQ(ierr); 65 ierr = MatDestroy(&ptap->P_loc);CHKERRQ(ierr); 66 ierr = MatDestroy(&ptap->P_oth);CHKERRQ(ierr); 67 ierr = PetscFree(ptap->api);CHKERRQ(ierr); 68 ierr = PetscFree(ptap->apj);CHKERRQ(ierr); 69 ierr = PetscFree(ptap->apa);CHKERRQ(ierr); 70 ierr = ptap->destroy(A);CHKERRQ(ierr); 71 ierr = PetscFree(ptap);CHKERRQ(ierr); 72 PetscFunctionReturn(0); 73 } 74 75 #undef __FUNCT__ 76 #define __FUNCT__ "MatDestroy_MPIAIJ_MatMatMult_32" 77 PetscErrorCode MatDestroy_MPIAIJ_MatMatMult_32(Mat A) 78 { 79 PetscErrorCode ierr; 80 PetscContainer container; 81 Mat_MatMatMultMPI *mult=PETSC_NULL; 82 83 PetscFunctionBegin; 84 ierr = PetscObjectQuery((PetscObject)A,"Mat_MatMatMultMPI",(PetscObject *)&container);CHKERRQ(ierr); 85 if (!container) SETERRQ(((PetscObject)A)->comm,PETSC_ERR_PLIB,"Container does not exist"); 86 ierr = PetscContainerGetPointer(container,(void **)&mult);CHKERRQ(ierr); 87 A->ops->destroy = mult->destroy; 88 A->ops->duplicate = mult->duplicate; 89 if (A->ops->destroy) { 90 ierr = (*A->ops->destroy)(A);CHKERRQ(ierr); 91 } 92 ierr = PetscObjectCompose((PetscObject)A,"Mat_MatMatMultMPI",0);CHKERRQ(ierr); 93 PetscFunctionReturn(0); 94 } 95 96 #undef __FUNCT__ 97 #define __FUNCT__ "MatDuplicate_MPIAIJ_MatMatMult_32" 98 PetscErrorCode MatDuplicate_MPIAIJ_MatMatMult_32(Mat A, MatDuplicateOption op, Mat *M) 99 { 100 PetscErrorCode ierr; 101 Mat_MatMatMultMPI *mult; 102 PetscContainer container; 103 104 PetscFunctionBegin; 105 ierr = PetscObjectQuery((PetscObject)A,"Mat_MatMatMultMPI",(PetscObject *)&container);CHKERRQ(ierr); 106 if (!container) SETERRQ(((PetscObject)A)->comm,PETSC_ERR_PLIB,"Container does not exist"); 107 ierr = PetscContainerGetPointer(container,(void **)&mult);CHKERRQ(ierr); 108 /* Note: the container is not duplicated, because it requires deep copying of 109 several large data sets (see PetscContainerDestroy_Mat_MatMatMultMPI()). 110 These data sets are only used for repeated calling of MatMatMultNumeric(). 111 *M is unlikely being used in this way. Thus we create *M with pure mpiaij format */ 112 ierr = (*mult->duplicate)(A,op,M);CHKERRQ(ierr); 113 (*M)->ops->destroy = mult->destroy; /* = MatDestroy_MPIAIJ, *M doesn't duplicate A's container! */ 114 (*M)->ops->duplicate = mult->duplicate; /* = MatDuplicate_MPIAIJ */ 115 PetscFunctionReturn(0); 116 } 117 118 #undef __FUNCT__ 119 #define __FUNCT__ "MatDuplicate_MPIAIJ_MatMatMult" 120 PetscErrorCode MatDuplicate_MPIAIJ_MatMatMult(Mat A, MatDuplicateOption op, Mat *M) 121 { 122 PetscErrorCode ierr; 123 Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data; 124 Mat_PtAPMPI *ptap=a->ptap; 125 126 PetscFunctionBegin; 127 ierr = (*ptap->duplicate)(A,op,M);CHKERRQ(ierr); 128 (*M)->ops->destroy = ptap->destroy; /* = MatDestroy_MPIAIJ, *M doesn't duplicate A's special structure! */ 129 (*M)->ops->duplicate = ptap->duplicate; /* = MatDuplicate_MPIAIJ */ 130 PetscFunctionReturn(0); 131 } 132 133 #undef __FUNCT__ 134 #define __FUNCT__ "MatMatMultNumeric_MPIAIJ_MPIAIJ" 135 PetscErrorCode MatMatMultNumeric_MPIAIJ_MPIAIJ(Mat A,Mat P,Mat C) 136 { 137 PetscErrorCode ierr; 138 Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data,*c=(Mat_MPIAIJ*)C->data; 139 Mat_SeqAIJ *ad=(Mat_SeqAIJ*)(a->A)->data,*ao=(Mat_SeqAIJ*)(a->B)->data; 140 Mat_SeqAIJ *cd=(Mat_SeqAIJ*)(c->A)->data,*co=(Mat_SeqAIJ*)(c->B)->data; 141 PetscInt *adi=ad->i,*adj,*aoi=ao->i,*aoj; 142 PetscScalar *ada,*aoa,*cda=cd->a,*coa=co->a; 143 Mat_SeqAIJ *p_loc,*p_oth; 144 PetscInt *pi_loc,*pj_loc,*pi_oth,*pj_oth,*pj; 145 PetscScalar *pa_loc,*pa_oth,*pa,*apa,valtmp,*ca; 146 PetscInt cm=C->rmap->n,anz,pnz; 147 Mat_PtAPMPI *ptap=c->ptap; 148 PetscInt *api,*apj,*apJ,i,j,k,row; 149 PetscInt cstart=C->cmap->rstart; 150 PetscInt cdnz,conz,k0,k1; 151 #if defined(DEBUG_MATMATMULT) 152 PetscMPIInt rank=a->rank; 153 #endif 154 155 PetscFunctionBegin; 156 #if defined(DEBUG_MATMATMULT) 157 if (!rank) ierr = PetscPrintf(PETSC_COMM_SELF,"[%d] call MatMatMultNumeric_MPIAIJ_MPIAIJ()...\n",rank); 158 #endif 159 160 /* 1) get P_oth = ptap->P_oth and P_loc = ptap->P_loc */ 161 /*-----------------------------------------------------*/ 162 /* update numerical values of P_oth and P_loc */ 163 ierr = MatGetBrowsOfAoCols_MPIAIJ(A,P,MAT_REUSE_MATRIX,&ptap->startsj,&ptap->startsj_r,&ptap->bufa,&ptap->P_oth);CHKERRQ(ierr); 164 ierr = MatMPIAIJGetLocalMat(P,MAT_REUSE_MATRIX,&ptap->P_loc);CHKERRQ(ierr); 165 #if defined(DEBUG_MATMATMULT) 166 if (!rank) ierr = PetscPrintf(PETSC_COMM_SELF,"[%d] got P_oth and P_loc...\n",rank); 167 #endif 168 169 /* 2) compute numeric C_loc = A_loc*P = Ad*P_loc + Ao*P_oth */ 170 /*----------------------------------------------------------*/ 171 /* get data from symbolic products */ 172 p_loc = (Mat_SeqAIJ*)(ptap->P_loc)->data; 173 p_oth = (Mat_SeqAIJ*)(ptap->P_oth)->data; 174 pi_loc=p_loc->i; pj_loc=p_loc->j; pa_loc=p_loc->a; 175 pi_oth=p_oth->i; pj_oth=p_oth->j; pa_oth=p_oth->a; 176 177 /* get apa for storing dense row A[i,:]*P */ 178 apa = ptap->apa; 179 180 api = ptap->api; 181 apj = ptap->apj; 182 for (i=0; i<cm; i++) { 183 /* diagonal portion of A */ 184 anz = adi[i+1] - adi[i]; 185 adj = ad->j + adi[i]; 186 ada = ad->a + adi[i]; 187 for (j=0; j<anz; j++) { 188 row = adj[j]; 189 pnz = pi_loc[row+1] - pi_loc[row]; 190 pj = pj_loc + pi_loc[row]; 191 pa = pa_loc + pi_loc[row]; 192 193 /* perform dense axpy */ 194 valtmp = ada[j]; 195 for (k=0; k<pnz; k++){ 196 apa[pj[k]] += valtmp*pa[k]; 197 } 198 ierr = PetscLogFlops(2.0*pnz);CHKERRQ(ierr); 199 } 200 201 /* off-diagonal portion of A */ 202 anz = aoi[i+1] - aoi[i]; 203 aoj = ao->j + aoi[i]; 204 aoa = ao->a + aoi[i]; 205 for (j=0; j<anz; j++) { 206 row = aoj[j]; 207 pnz = pi_oth[row+1] - pi_oth[row]; 208 pj = pj_oth + pi_oth[row]; 209 pa = pa_oth + pi_oth[row]; 210 211 /* perform dense axpy */ 212 valtmp = aoa[j]; 213 for (k=0; k<pnz; k++){ 214 apa[pj[k]] += valtmp*pa[k]; 215 } 216 ierr = PetscLogFlops(2.0*pnz);CHKERRQ(ierr); 217 } 218 219 /* set values in C */ 220 apJ = apj + api[i]; 221 cdnz = cd->i[i+1] - cd->i[i]; 222 conz = co->i[i+1] - co->i[i]; 223 224 /* 1st off-diagoanl part of C */ 225 ca = coa + co->i[i]; 226 k = 0; 227 for (k0=0; k0<conz; k0++){ 228 if (apJ[k] >= cstart) break; 229 ca[k0] = apa[apJ[k]]; 230 apa[apJ[k]] = 0.0; 231 k++; 232 } 233 234 /* diagonal part of C */ 235 ca = cda + cd->i[i]; 236 for (k1=0; k1<cdnz; k1++){ 237 ca[k1] = apa[apJ[k]]; 238 apa[apJ[k]] = 0.0; 239 k++; 240 } 241 242 /* 2nd off-diagoanl part of C */ 243 ca = coa + co->i[i]; 244 for (; k0<conz; k0++){ 245 ca[k0] = apa[apJ[k]]; 246 apa[apJ[k]] = 0.0; 247 k++; 248 } 249 } 250 ierr = MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 251 ierr = MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 252 PetscFunctionReturn(0); 253 } 254 255 #undef __FUNCT__ 256 #define __FUNCT__ "MatMatMultSymbolic_MPIAIJ_MPIAIJ" 257 PetscErrorCode MatMatMultSymbolic_MPIAIJ_MPIAIJ(Mat A,Mat P,PetscReal fill,Mat *C) 258 { 259 PetscErrorCode ierr; 260 MPI_Comm comm=((PetscObject)A)->comm; 261 Mat Cmpi; 262 Mat_PtAPMPI *ptap; 263 PetscFreeSpaceList free_space=PETSC_NULL,current_space=PETSC_NULL; 264 Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data,*c; 265 Mat_SeqAIJ *ad=(Mat_SeqAIJ*)(a->A)->data,*ao=(Mat_SeqAIJ*)(a->B)->data,*p_loc,*p_oth; 266 PetscInt *pi_loc,*pj_loc,*pi_oth,*pj_oth,*dnz,*onz; 267 PetscInt *adi=ad->i,*adj=ad->j,*aoi=ao->i,*aoj=ao->j,rstart=A->rmap->rstart; 268 PetscInt *lnk,i,pnz,row,*api,*apj,*Jptr,apnz,nspacedouble=0,j,nzi; 269 PetscInt am=A->rmap->n,pN=P->cmap->N,pn=P->cmap->n,pm=P->rmap->n; 270 PetscBT lnkbt; 271 PetscScalar *apa; 272 PetscReal afill; 273 PetscBool scalable=PETSC_FALSE; 274 PetscInt nlnk_max,armax,prmax; 275 #if defined(DEBUG_MATMATMULT) 276 PetscMPIInt rank=a->rank; 277 #endif 278 279 PetscFunctionBegin; 280 if (A->cmap->rstart != P->rmap->rstart || A->cmap->rend != P->rmap->rend){ 281 SETERRQ4(comm,PETSC_ERR_ARG_SIZ,"Matrix local dimensions are incompatible, (%D, %D) != (%D,%D)",A->cmap->rstart,A->cmap->rend,P->rmap->rstart,P->rmap->rend); 282 } 283 284 ierr = PetscObjectOptionsBegin((PetscObject)A);CHKERRQ(ierr); 285 286 ierr = PetscOptionsBool("-matmatmult_32","Use a scalable but slower C=A*B","",scalable,&scalable,PETSC_NULL);CHKERRQ(ierr); 287 if (scalable){ 288 ierr = MatMatMultSymbolic_MPIAIJ_MPIAIJ_Scalable_32(A,P,fill,C);;CHKERRQ(ierr); 289 PetscFunctionReturn(0); 290 } 291 ierr = PetscOptionsBool("-matmatmult_scalable","Use a scalable but slower C=A*B","",scalable,&scalable,PETSC_NULL);CHKERRQ(ierr); 292 if (scalable){ 293 ierr = MatMatMultSymbolic_MPIAIJ_MPIAIJ_Scalable(A,P,fill,C);;CHKERRQ(ierr); 294 PetscFunctionReturn(0); 295 } 296 ierr = PetscOptionsEnd();CHKERRQ(ierr); 297 298 #if defined(DEBUG_MATMATMULT) 299 if (!rank) ierr = PetscPrintf(PETSC_COMM_SELF,"[%d] call MatMatMultSymbolic_MPIAIJ_MPIAIJ()...\n",rank); 300 #endif 301 302 /* create struct Mat_PtAPMPI and attached it to C later */ 303 ierr = PetscNew(Mat_PtAPMPI,&ptap);CHKERRQ(ierr); 304 305 /* get P_oth by taking rows of P (= non-zero cols of local A) from other processors */ 306 ierr = MatGetBrowsOfAoCols_MPIAIJ(A,P,MAT_INITIAL_MATRIX,&ptap->startsj,&ptap->startsj_r,&ptap->bufa,&ptap->P_oth);CHKERRQ(ierr); 307 #if defined(DEBUG_MATMATMULT) 308 if (!rank) ierr = PetscPrintf(PETSC_COMM_SELF,"[%d] P_oth is done...\n",rank); 309 #endif 310 /* get P_loc by taking all local rows of P */ 311 ierr = MatMPIAIJGetLocalMat(P,MAT_INITIAL_MATRIX,&ptap->P_loc);CHKERRQ(ierr); 312 313 p_loc = (Mat_SeqAIJ*)(ptap->P_loc)->data; 314 p_oth = (Mat_SeqAIJ*)(ptap->P_oth)->data; 315 pi_loc = p_loc->i; pj_loc = p_loc->j; 316 pi_oth = p_oth->i; pj_oth = p_oth->j; 317 318 #if defined(DEBUG_MATMATMULT) 319 if (!rank) ierr = PetscPrintf(PETSC_COMM_SELF,"[%d] P_loc is done, Annz %D * P_locnnz %D = %D...\n",rank,ad->i[am]+ao->i[am],p_loc->rmax,(ad->i[am]+ao->i[am])*p_loc->rmax); 320 #endif 321 322 /* first, compute symbolic AP = A_loc*P = A_diag*P_loc + A_off*P_oth */ 323 /*-------------------------------------------------------------------*/ 324 ierr = PetscMalloc((am+2)*sizeof(PetscInt),&api);CHKERRQ(ierr); 325 ptap->api = api; 326 api[0] = 0; 327 328 /* create and initialize a linked list */ 329 armax = ad->rmax+ao->rmax; 330 prmax = PetscMax(p_loc->rmax,p_oth->rmax); 331 nlnk_max = armax*prmax; 332 if (!nlnk_max || nlnk_max > pN) nlnk_max = pN; 333 #if defined(DEBUG_MATMATMULT) 334 if (!rank) ierr = PetscPrintf(PETSC_COMM_SELF,"[%d] pN %d; nlnk_max %d; Armax %d+%d=%d; Prmax Max(%d,%d)=%d\n",rank,pN,nlnk_max,ad->rmax,ao->rmax,armax,p_loc->rmax,p_oth->rmax,prmax); 335 #endif 336 ierr = PetscLLCondensedCreate(nlnk_max,pN,&lnk,&lnkbt);CHKERRQ(ierr); 337 338 /* Initial FreeSpace size is fill*(nnz(A)+nnz(P)) */ 339 ierr = PetscFreeSpaceGet((PetscInt)(fill*(adi[am]+aoi[am]+pi_loc[pm])),&free_space);CHKERRQ(ierr); 340 current_space = free_space; 341 342 ierr = MatPreallocateInitialize(comm,am,pn,dnz,onz);CHKERRQ(ierr); 343 for (i=0; i<am; i++) { 344 apnz = 0; 345 /* diagonal portion of A */ 346 nzi = adi[i+1] - adi[i]; 347 for (j=0; j<nzi; j++){ 348 row = *adj++; 349 pnz = pi_loc[row+1] - pi_loc[row]; 350 Jptr = pj_loc + pi_loc[row]; 351 /* add non-zero cols of P into the sorted linked list lnk */ 352 ierr = PetscLLCondensedAddSorted(pnz,Jptr,lnk,lnkbt);CHKERRQ(ierr); 353 } 354 /* off-diagonal portion of A */ 355 nzi = aoi[i+1] - aoi[i]; 356 for (j=0; j<nzi; j++){ 357 row = *aoj++; 358 pnz = pi_oth[row+1] - pi_oth[row]; 359 Jptr = pj_oth + pi_oth[row]; 360 ierr = PetscLLCondensedAddSorted(pnz,Jptr,lnk,lnkbt);CHKERRQ(ierr); 361 } 362 363 apnz = lnk[0]; 364 api[i+1] = api[i] + apnz; 365 366 /* if free space is not available, double the total space in the list */ 367 if (current_space->local_remaining<apnz) { 368 ierr = PetscFreeSpaceGet(apnz+current_space->total_array_size,¤t_space);CHKERRQ(ierr); 369 nspacedouble++; 370 } 371 372 /* Copy data into free space, then initialize lnk */ 373 ierr = PetscLLCondensedClean(pN,apnz,current_space->array,lnk,lnkbt);CHKERRQ(ierr); 374 ierr = MatPreallocateSet(i+rstart,apnz,current_space->array,dnz,onz);CHKERRQ(ierr); 375 current_space->array += apnz; 376 current_space->local_used += apnz; 377 current_space->local_remaining -= apnz; 378 } 379 380 /* Allocate space for apj, initialize apj, and */ 381 /* destroy list of free space and other temporary array(s) */ 382 ierr = PetscMalloc((api[am]+1)*sizeof(PetscInt),&ptap->apj);CHKERRQ(ierr); 383 apj = ptap->apj; 384 ierr = PetscFreeSpaceContiguous(&free_space,ptap->apj);CHKERRQ(ierr); 385 ierr = PetscLLDestroy(lnk,lnkbt);CHKERRQ(ierr); 386 #if defined(DEBUG_MATMATMULT) 387 if (!rank) ierr = PetscPrintf(PETSC_COMM_SELF,"[%d] AP is done...\n",rank); 388 #endif 389 390 /* malloc apa to store dense row A[i,:]*P */ 391 ierr = PetscMalloc(pN*sizeof(PetscScalar),&apa);CHKERRQ(ierr); 392 ierr = PetscMemzero(apa,pN*sizeof(PetscScalar));CHKERRQ(ierr); 393 ptap->apa = apa; 394 #if defined(DEBUG_MATMATMULT) 395 if (!rank) ierr = PetscPrintf(PETSC_COMM_SELF,"[%d] Malloc apa pN %D is done...\n",rank,pN); 396 #endif 397 398 /* create and assemble symbolic parallel matrix Cmpi */ 399 /*----------------------------------------------------*/ 400 ierr = MatCreate(comm,&Cmpi);CHKERRQ(ierr); 401 ierr = MatSetSizes(Cmpi,am,pn,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 402 ierr = MatSetType(Cmpi,MATMPIAIJ);CHKERRQ(ierr); 403 ierr = MatMPIAIJSetPreallocation(Cmpi,0,dnz,0,onz);CHKERRQ(ierr); 404 ierr = MatPreallocateFinalize(dnz,onz);CHKERRQ(ierr); 405 ierr = MatSetBlockSize(Cmpi,1);CHKERRQ(ierr); 406 for (i=0; i<am; i++){ 407 row = i + rstart; 408 apnz = api[i+1] - api[i]; 409 ierr = MatSetValues(Cmpi,1,&row,apnz,apj,apa,INSERT_VALUES);CHKERRQ(ierr); 410 apj += apnz; 411 } 412 ierr = MatAssemblyBegin(Cmpi,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 413 ierr = MatAssemblyEnd(Cmpi,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 414 415 ptap->destroy = Cmpi->ops->destroy; 416 ptap->duplicate = Cmpi->ops->duplicate; 417 Cmpi->ops->matmultnumeric = MatMatMultNumeric_MPIAIJ_MPIAIJ; 418 Cmpi->ops->destroy = MatDestroy_MPIAIJ_MatMatMult; 419 Cmpi->ops->duplicate = MatDuplicate_MPIAIJ_MatMatMult; 420 421 /* attach the supporting struct to Cmpi for reuse */ 422 c = (Mat_MPIAIJ*)Cmpi->data; 423 c->ptap = ptap; 424 425 *C = Cmpi; 426 427 /* set MatInfo */ 428 afill = (PetscReal)api[am]/(adi[am]+aoi[am]+pi_loc[pm]) + 1.e-5; 429 if (afill < 1.0) afill = 1.0; 430 Cmpi->info.mallocs = nspacedouble; 431 Cmpi->info.fill_ratio_given = fill; 432 Cmpi->info.fill_ratio_needed = afill; 433 434 #if defined(PETSC_USE_INFO) 435 if (api[am]) { 436 ierr = PetscInfo3(Cmpi,"Reallocs %D; Fill ratio: given %G needed %G.\n",nspacedouble,fill,afill);CHKERRQ(ierr); 437 ierr = PetscInfo1(Cmpi,"Use MatMatMult(A,B,MatReuse,%G,&C) for best performance.;\n",afill);CHKERRQ(ierr); 438 } else { 439 ierr = PetscInfo(Cmpi,"Empty matrix product\n");CHKERRQ(ierr); 440 } 441 #endif 442 PetscFunctionReturn(0); 443 } 444 445 /* implementation used in PETSc-3.2 */ 446 /* This routine is called ONLY in the case of reusing previously computed symbolic C */ 447 #undef __FUNCT__ 448 #define __FUNCT__ "MatMatMultNumeric_MPIAIJ_MPIAIJ_Scalable_32" 449 PetscErrorCode MatMatMultNumeric_MPIAIJ_MPIAIJ_Scalable_32(Mat A,Mat B,Mat C) 450 { 451 PetscErrorCode ierr; 452 Mat *seq; 453 Mat_MatMatMultMPI *mult; 454 PetscContainer container; 455 456 PetscFunctionBegin; 457 ierr = PetscObjectQuery((PetscObject)C,"Mat_MatMatMultMPI",(PetscObject *)&container);CHKERRQ(ierr); 458 if (!container) SETERRQ(((PetscObject)A)->comm,PETSC_ERR_PLIB,"Container does not exist"); 459 ierr = PetscContainerGetPointer(container,(void **)&mult);CHKERRQ(ierr); 460 461 if (mult->skipNumeric){ /* first numeric product is done during symbolic product */ 462 mult->skipNumeric = PETSC_FALSE; 463 PetscFunctionReturn(0); 464 } 465 #if defined(DEBUG_MATMATMULT) 466 PetscMPIInt rank; 467 MPI_Comm comm = ((PetscObject)C)->comm; 468 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 469 ierr = MPI_Barrier(comm);CHKERRQ(ierr); 470 if (!rank) ierr = PetscPrintf(PETSC_COMM_SELF,"[%d] call MatMatMultNumeric_MPIAIJ_MPIAIJ_Scalable_32()...\n",rank); 471 #endif 472 473 seq = &mult->B_seq; 474 ierr = MatGetSubMatrices(B,1,&mult->isrowb,&mult->iscolb,MAT_REUSE_MATRIX,&seq);CHKERRQ(ierr); 475 mult->B_seq = *seq; 476 477 seq = &mult->A_loc; 478 ierr = MatGetSubMatrices(A,1,&mult->isrowa,&mult->isrowb,MAT_REUSE_MATRIX,&seq);CHKERRQ(ierr); 479 mult->A_loc = *seq; 480 481 ierr = MatMatMultNumeric_SeqAIJ_SeqAIJ_Scalable(mult->A_loc,mult->B_seq,mult->C_seq);CHKERRQ(ierr); 482 483 ierr = PetscObjectReference((PetscObject)mult->C_seq);CHKERRQ(ierr); 484 ierr = MatMerge(((PetscObject)A)->comm,mult->C_seq,B->cmap->n,MAT_REUSE_MATRIX,&C);CHKERRQ(ierr); 485 PetscFunctionReturn(0); 486 } 487 488 /* numeric product is computed as well */ 489 #undef __FUNCT__ 490 #define __FUNCT__ "MatMatMultSymbolic_MPIAIJ_MPIAIJ_Scalable_32" 491 PetscErrorCode MatMatMultSymbolic_MPIAIJ_MPIAIJ_Scalable_32(Mat A,Mat B,PetscReal fill,Mat *C) 492 { 493 PetscErrorCode ierr; 494 Mat_MatMatMultMPI *mult; 495 PetscContainer container; 496 Mat AB,*seq; 497 Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data; 498 PetscInt *idx,i,start,ncols,nzA,nzB,*cmap,imark; 499 #if defined(DEBUG_MATMATMULT) 500 MPI_Comm comm = ((PetscObject)A)->comm; 501 PetscMPIInt rank=a->rank; 502 #endif 503 504 PetscFunctionBegin; 505 #if defined(DEBUG_MATMATMULT) 506 if (!rank) ierr = PetscPrintf(PETSC_COMM_SELF,"[%d] call MatMatMultSymbolic_MPIAIJ_MPIAIJ_Scalable_32()...\n",rank); 507 #endif 508 if (A->cmap->rstart != B->rmap->rstart || A->cmap->rend != B->rmap->rend){ 509 SETERRQ4(((PetscObject)A)->comm,PETSC_ERR_ARG_SIZ,"Matrix local dimensions are incompatible, (%D, %D) != (%D,%D)",A->cmap->rstart,A->cmap->rend,B->rmap->rstart,B->rmap->rend); 510 } 511 512 ierr = PetscNew(Mat_MatMatMultMPI,&mult);CHKERRQ(ierr); 513 514 /* get isrowb: nonzero col of A */ 515 start = A->cmap->rstart; 516 cmap = a->garray; 517 nzA = a->A->cmap->n; 518 nzB = a->B->cmap->n; 519 ierr = PetscMalloc((nzA+nzB)*sizeof(PetscInt), &idx);CHKERRQ(ierr); 520 ncols = 0; 521 for (i=0; i<nzB; i++) { /* row < local row index */ 522 if (cmap[i] < start) idx[ncols++] = cmap[i]; 523 else break; 524 } 525 imark = i; 526 for (i=0; i<nzA; i++) idx[ncols++] = start + i; /* local rows */ 527 for (i=imark; i<nzB; i++) idx[ncols++] = cmap[i]; /* row > local row index */ 528 ierr = ISCreateGeneral(PETSC_COMM_SELF,ncols,idx,PETSC_OWN_POINTER,&mult->isrowb);CHKERRQ(ierr); 529 ierr = ISCreateStride(PETSC_COMM_SELF,B->cmap->N,0,1,&mult->iscolb);CHKERRQ(ierr); 530 531 /* get isrowa: all local rows of A */ 532 ierr = ISCreateStride(PETSC_COMM_SELF,A->rmap->n,A->rmap->rstart,1,&mult->isrowa);CHKERRQ(ierr); 533 534 /* Below should go to MatMatMultNumeric_MPIAIJ_MPIAIJ() - How to generate C there? */ 535 /* create a seq matrix B_seq = submatrix of B by taking rows of B that equal to nonzero col of A */ 536 ierr = MatGetSubMatrices(B,1,&mult->isrowb,&mult->iscolb,MAT_INITIAL_MATRIX,&seq);CHKERRQ(ierr); 537 mult->B_seq = *seq; 538 ierr = PetscFree(seq);CHKERRQ(ierr); 539 #if defined(DEBUG_MATMATMULT) 540 if (!rank) ierr = PetscPrintf(PETSC_COMM_SELF,"[%d] B_seq is done...\n",rank); 541 #endif 542 543 /* create a seq matrix A_seq = submatrix of A by taking all local rows of A */ 544 ierr = MatGetSubMatrices(A,1,&mult->isrowa,&mult->isrowb,MAT_INITIAL_MATRIX,&seq);CHKERRQ(ierr); 545 mult->A_loc = *seq; 546 ierr = PetscFree(seq);CHKERRQ(ierr); 547 #if defined(DEBUG_MATMATMULT) 548 if (!rank) ierr = PetscPrintf(PETSC_COMM_SELF,"[%d] A_loc is done...\n",rank); 549 #endif 550 551 /* compute C_seq = A_seq * B_seq */ 552 ierr = MatMatMultSymbolic_SeqAIJ_SeqAIJ_Scalable(mult->A_loc,mult->B_seq,fill,&mult->C_seq);CHKERRQ(ierr); 553 #if defined(DEBUG_MATMATMULT) 554 ierr = MPI_Barrier(comm);CHKERRQ(ierr); 555 if (!rank) ierr = PetscPrintf(PETSC_COMM_SELF,"[%d] C_seq Symbolic is done...\n",rank); 556 #endif 557 ierr = MatMatMultNumeric_SeqAIJ_SeqAIJ_Scalable(mult->A_loc,mult->B_seq,mult->C_seq);CHKERRQ(ierr); 558 #if defined(DEBUG_MATMATMULT) 559 if (!rank) ierr = PetscPrintf(PETSC_COMM_SELF,"[%d] C_seq Numeric is done...\n",rank); 560 #endif 561 562 /* create mpi matrix C by concatinating C_seq */ 563 ierr = PetscObjectReference((PetscObject)mult->C_seq);CHKERRQ(ierr); /* prevent C_seq being destroyed by MatMerge() */ 564 ierr = MatMergeSymbolic(((PetscObject)A)->comm,mult->C_seq,B->cmap->n,&AB);CHKERRQ(ierr); 565 ierr = MatMergeNumeric(((PetscObject)A)->comm,mult->C_seq,B->cmap->n,AB);CHKERRQ(ierr); 566 #if defined(DEBUG_MATMATMULT) 567 if (!rank) ierr = PetscPrintf(PETSC_COMM_SELF,"[%d] Merge is done...\n",rank); 568 #endif 569 570 /* attach the supporting struct to C for reuse of symbolic C */ 571 ierr = PetscContainerCreate(PETSC_COMM_SELF,&container);CHKERRQ(ierr); 572 ierr = PetscContainerSetPointer(container,mult);CHKERRQ(ierr); 573 ierr = PetscContainerSetUserDestroy(container,PetscContainerDestroy_Mat_MatMatMultMPI);CHKERRQ(ierr); 574 ierr = PetscObjectCompose((PetscObject)AB,"Mat_MatMatMultMPI",(PetscObject)container);CHKERRQ(ierr); 575 ierr = PetscContainerDestroy(&container);CHKERRQ(ierr); 576 mult->skipNumeric = PETSC_TRUE; /* a numeric product is done here */ 577 mult->destroy = AB->ops->destroy; 578 mult->duplicate = AB->ops->duplicate; 579 AB->ops->matmultnumeric = MatMatMultNumeric_MPIAIJ_MPIAIJ_Scalable_32; 580 AB->ops->destroy = MatDestroy_MPIAIJ_MatMatMult_32; 581 AB->ops->duplicate = MatDuplicate_MPIAIJ_MatMatMult_32; 582 AB->ops->matmult = MatMatMult_MPIAIJ_MPIAIJ; 583 *C = AB; 584 PetscFunctionReturn(0); 585 } 586 587 #undef __FUNCT__ 588 #define __FUNCT__ "MatMatMult_MPIAIJ_MPIDense" 589 PetscErrorCode MatMatMult_MPIAIJ_MPIDense(Mat A,Mat B,MatReuse scall,PetscReal fill,Mat *C) 590 { 591 PetscErrorCode ierr; 592 593 PetscFunctionBegin; 594 if (scall == MAT_INITIAL_MATRIX){ 595 ierr = MatMatMultSymbolic_MPIAIJ_MPIDense(A,B,fill,C);CHKERRQ(ierr); 596 } 597 ierr = MatMatMultNumeric_MPIAIJ_MPIDense(A,B,*C);CHKERRQ(ierr); 598 PetscFunctionReturn(0); 599 } 600 601 typedef struct { 602 Mat workB; 603 PetscScalar *rvalues,*svalues; 604 MPI_Request *rwaits,*swaits; 605 } MPIAIJ_MPIDense; 606 607 #undef __FUNCT__ 608 #define __FUNCT__ "MPIAIJ_MPIDenseDestroy" 609 PetscErrorCode MPIAIJ_MPIDenseDestroy(void *ctx) 610 { 611 MPIAIJ_MPIDense *contents = (MPIAIJ_MPIDense*) ctx; 612 PetscErrorCode ierr; 613 614 PetscFunctionBegin; 615 ierr = MatDestroy(&contents->workB);CHKERRQ(ierr); 616 ierr = PetscFree4(contents->rvalues,contents->svalues,contents->rwaits,contents->swaits);CHKERRQ(ierr); 617 ierr = PetscFree(contents);CHKERRQ(ierr); 618 PetscFunctionReturn(0); 619 } 620 621 #undef __FUNCT__ 622 #define __FUNCT__ "MatMatMultSymbolic_MPIAIJ_MPIDense" 623 PetscErrorCode MatMatMultSymbolic_MPIAIJ_MPIDense(Mat A,Mat B,PetscReal fill,Mat *C) 624 { 625 PetscErrorCode ierr; 626 Mat_MPIAIJ *aij = (Mat_MPIAIJ*) A->data; 627 PetscInt nz = aij->B->cmap->n; 628 PetscContainer container; 629 MPIAIJ_MPIDense *contents; 630 VecScatter ctx = aij->Mvctx; 631 VecScatter_MPI_General *from = (VecScatter_MPI_General*) ctx->fromdata; 632 VecScatter_MPI_General *to = ( VecScatter_MPI_General*) ctx->todata; 633 PetscInt m=A->rmap->n,n=B->cmap->n; 634 635 PetscFunctionBegin; 636 ierr = MatCreate(((PetscObject)B)->comm,C);CHKERRQ(ierr); 637 ierr = MatSetSizes(*C,m,n,A->rmap->N,B->cmap->N);CHKERRQ(ierr); 638 ierr = MatSetType(*C,MATMPIDENSE);CHKERRQ(ierr); 639 ierr = MatAssemblyBegin(*C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 640 ierr = MatAssemblyEnd(*C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 641 (*C)->ops->matmult = MatMatMult_MPIAIJ_MPIDense; 642 643 ierr = PetscNew(MPIAIJ_MPIDense,&contents);CHKERRQ(ierr); 644 /* Create work matrix used to store off processor rows of B needed for local product */ 645 ierr = MatCreateSeqDense(PETSC_COMM_SELF,nz,B->cmap->N,PETSC_NULL,&contents->workB);CHKERRQ(ierr); 646 /* Create work arrays needed */ 647 ierr = PetscMalloc4(B->cmap->N*from->starts[from->n],PetscScalar,&contents->rvalues, 648 B->cmap->N*to->starts[to->n],PetscScalar,&contents->svalues, 649 from->n,MPI_Request,&contents->rwaits, 650 to->n,MPI_Request,&contents->swaits);CHKERRQ(ierr); 651 652 ierr = PetscContainerCreate(((PetscObject)A)->comm,&container);CHKERRQ(ierr); 653 ierr = PetscContainerSetPointer(container,contents);CHKERRQ(ierr); 654 ierr = PetscContainerSetUserDestroy(container,MPIAIJ_MPIDenseDestroy);CHKERRQ(ierr); 655 ierr = PetscObjectCompose((PetscObject)(*C),"workB",(PetscObject)container);CHKERRQ(ierr); 656 ierr = PetscContainerDestroy(&container);CHKERRQ(ierr); 657 PetscFunctionReturn(0); 658 } 659 660 #undef __FUNCT__ 661 #define __FUNCT__ "MatMPIDenseScatter" 662 /* 663 Performs an efficient scatter on the rows of B needed by this process; this is 664 a modification of the VecScatterBegin_() routines. 665 */ 666 PetscErrorCode MatMPIDenseScatter(Mat A,Mat B,Mat C,Mat *outworkB) 667 { 668 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)A->data; 669 PetscErrorCode ierr; 670 PetscScalar *b,*w,*svalues,*rvalues; 671 VecScatter ctx = aij->Mvctx; 672 VecScatter_MPI_General *from = (VecScatter_MPI_General*) ctx->fromdata; 673 VecScatter_MPI_General *to = ( VecScatter_MPI_General*) ctx->todata; 674 PetscInt i,j,k; 675 PetscInt *sindices,*sstarts,*rindices,*rstarts; 676 PetscMPIInt *sprocs,*rprocs,nrecvs; 677 MPI_Request *swaits,*rwaits; 678 MPI_Comm comm = ((PetscObject)A)->comm; 679 PetscMPIInt tag = ((PetscObject)ctx)->tag,ncols = B->cmap->N, nrows = aij->B->cmap->n,imdex,nrowsB = B->rmap->n; 680 MPI_Status status; 681 MPIAIJ_MPIDense *contents; 682 PetscContainer container; 683 Mat workB; 684 685 PetscFunctionBegin; 686 ierr = PetscObjectQuery((PetscObject)C,"workB",(PetscObject*)&container);CHKERRQ(ierr); 687 if (!container) SETERRQ(comm,PETSC_ERR_PLIB,"Container does not exist"); 688 ierr = PetscContainerGetPointer(container,(void**)&contents);CHKERRQ(ierr); 689 690 workB = *outworkB = contents->workB; 691 if (nrows != workB->rmap->n) SETERRQ2(comm,PETSC_ERR_PLIB,"Number of rows of workB %D not equal to columns of aij->B %D",nrows,workB->cmap->n); 692 sindices = to->indices; 693 sstarts = to->starts; 694 sprocs = to->procs; 695 swaits = contents->swaits; 696 svalues = contents->svalues; 697 698 rindices = from->indices; 699 rstarts = from->starts; 700 rprocs = from->procs; 701 rwaits = contents->rwaits; 702 rvalues = contents->rvalues; 703 704 ierr = MatGetArray(B,&b);CHKERRQ(ierr); 705 ierr = MatGetArray(workB,&w);CHKERRQ(ierr); 706 707 for (i=0; i<from->n; i++) { 708 ierr = MPI_Irecv(rvalues+ncols*rstarts[i],ncols*(rstarts[i+1]-rstarts[i]),MPIU_SCALAR,rprocs[i],tag,comm,rwaits+i);CHKERRQ(ierr); 709 } 710 711 for (i=0; i<to->n; i++) { 712 /* pack a message at a time */ 713 CHKMEMQ; 714 for (j=0; j<sstarts[i+1]-sstarts[i]; j++){ 715 for (k=0; k<ncols; k++) { 716 svalues[ncols*(sstarts[i] + j) + k] = b[sindices[sstarts[i]+j] + nrowsB*k]; 717 } 718 } 719 CHKMEMQ; 720 ierr = MPI_Isend(svalues+ncols*sstarts[i],ncols*(sstarts[i+1]-sstarts[i]),MPIU_SCALAR,sprocs[i],tag,comm,swaits+i);CHKERRQ(ierr); 721 } 722 723 nrecvs = from->n; 724 while (nrecvs) { 725 ierr = MPI_Waitany(from->n,rwaits,&imdex,&status);CHKERRQ(ierr); 726 nrecvs--; 727 /* unpack a message at a time */ 728 CHKMEMQ; 729 for (j=0; j<rstarts[imdex+1]-rstarts[imdex]; j++){ 730 for (k=0; k<ncols; k++) { 731 w[rindices[rstarts[imdex]+j] + nrows*k] = rvalues[ncols*(rstarts[imdex] + j) + k]; 732 } 733 } 734 CHKMEMQ; 735 } 736 if (to->n) {ierr = MPI_Waitall(to->n,swaits,to->sstatus);CHKERRQ(ierr);} 737 738 ierr = MatRestoreArray(B,&b);CHKERRQ(ierr); 739 ierr = MatRestoreArray(workB,&w);CHKERRQ(ierr); 740 ierr = MatAssemblyBegin(workB,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 741 ierr = MatAssemblyEnd(workB,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 742 PetscFunctionReturn(0); 743 } 744 extern PetscErrorCode MatMatMultNumericAdd_SeqAIJ_SeqDense(Mat,Mat,Mat); 745 746 #undef __FUNCT__ 747 #define __FUNCT__ "MatMatMultNumeric_MPIAIJ_MPIDense" 748 PetscErrorCode MatMatMultNumeric_MPIAIJ_MPIDense(Mat A,Mat B,Mat C) 749 { 750 PetscErrorCode ierr; 751 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)A->data; 752 Mat_MPIDense *bdense = (Mat_MPIDense*)B->data; 753 Mat_MPIDense *cdense = (Mat_MPIDense*)C->data; 754 Mat workB; 755 756 PetscFunctionBegin; 757 758 /* diagonal block of A times all local rows of B*/ 759 ierr = MatMatMultNumeric_SeqAIJ_SeqDense(aij->A,bdense->A,cdense->A);CHKERRQ(ierr); 760 761 /* get off processor parts of B needed to complete the product */ 762 ierr = MatMPIDenseScatter(A,B,C,&workB);CHKERRQ(ierr); 763 764 /* off-diagonal block of A times nonlocal rows of B */ 765 ierr = MatMatMultNumericAdd_SeqAIJ_SeqDense(aij->B,workB,cdense->A);CHKERRQ(ierr); 766 ierr = MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 767 ierr = MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 768 PetscFunctionReturn(0); 769 } 770 771 #undef __FUNCT__ 772 #define __FUNCT__ "MatMatMultNumeric_MPIAIJ_MPIAIJ_Scalable" 773 PetscErrorCode MatMatMultNumeric_MPIAIJ_MPIAIJ_Scalable(Mat A,Mat P,Mat C) 774 { 775 PetscErrorCode ierr; 776 Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data,*c=(Mat_MPIAIJ*)C->data; 777 Mat_SeqAIJ *ad=(Mat_SeqAIJ*)(a->A)->data,*ao=(Mat_SeqAIJ*)(a->B)->data; 778 Mat_SeqAIJ *cd=(Mat_SeqAIJ*)(c->A)->data,*co=(Mat_SeqAIJ*)(c->B)->data; 779 PetscInt *adi=ad->i,*adj,*aoi=ao->i,*aoj; 780 PetscScalar *ada,*aoa,*cda=cd->a,*coa=co->a; 781 Mat_SeqAIJ *p_loc,*p_oth; 782 PetscInt *pi_loc,*pj_loc,*pi_oth,*pj_oth,*pj; 783 PetscScalar *pa_loc,*pa_oth,*pa,valtmp,*ca; 784 PetscInt cm=C->rmap->n,anz,pnz; 785 Mat_PtAPMPI *ptap=c->ptap; 786 PetscScalar *apa_sparse=ptap->apa; 787 PetscInt *api,*apj,*apJ,i,j,k,row; 788 PetscInt cstart=C->cmap->rstart; 789 PetscInt cdnz,conz,k0,k1,nextp; 790 #if defined(DEBUG_MATMATMULT) 791 PetscMPIInt rank=a->rank; 792 #endif 793 794 PetscFunctionBegin; 795 #if defined(DEBUG_MATMATMULT) 796 if (!rank) ierr = PetscPrintf(PETSC_COMM_SELF,"[%d] call MatMatMultNumeric_MPIAIJ_MPIAIJ_Scalable()...\n",rank); 797 #endif 798 799 /* 1) get P_oth = ptap->P_oth and P_loc = ptap->P_loc */ 800 /*-----------------------------------------------------*/ 801 /* update numerical values of P_oth and P_loc */ 802 ierr = MatGetBrowsOfAoCols_MPIAIJ(A,P,MAT_REUSE_MATRIX,&ptap->startsj,&ptap->startsj_r,&ptap->bufa,&ptap->P_oth);CHKERRQ(ierr); 803 ierr = MatMPIAIJGetLocalMat(P,MAT_REUSE_MATRIX,&ptap->P_loc);CHKERRQ(ierr); 804 #if defined(DEBUG_MATMATMULT) 805 if (!rank) ierr = PetscPrintf(PETSC_COMM_SELF,"[%d] got P_oth and P_loc...\n",rank); 806 #endif 807 808 /* 2) compute numeric C_loc = A_loc*P = Ad*P_loc + Ao*P_oth */ 809 /*----------------------------------------------------------*/ 810 /* get data from symbolic products */ 811 p_loc = (Mat_SeqAIJ*)(ptap->P_loc)->data; 812 p_oth = (Mat_SeqAIJ*)(ptap->P_oth)->data; 813 pi_loc=p_loc->i; pj_loc=p_loc->j; pa_loc=p_loc->a; 814 pi_oth=p_oth->i; pj_oth=p_oth->j; pa_oth=p_oth->a; 815 816 api = ptap->api; 817 apj = ptap->apj; 818 for (i=0; i<cm; i++) { 819 apJ = apj + api[i]; 820 821 /* diagonal portion of A */ 822 anz = adi[i+1] - adi[i]; 823 adj = ad->j + adi[i]; 824 ada = ad->a + adi[i]; 825 for (j=0; j<anz; j++) { 826 row = adj[j]; 827 pnz = pi_loc[row+1] - pi_loc[row]; 828 pj = pj_loc + pi_loc[row]; 829 pa = pa_loc + pi_loc[row]; 830 /* perform sparse axpy */ 831 valtmp = ada[j]; 832 nextp = 0; 833 for (k=0; nextp<pnz; k++) { 834 if (apJ[k] == pj[nextp]) { /* column of AP == column of P */ 835 apa_sparse[k] += valtmp*pa[nextp++]; 836 } 837 } 838 ierr = PetscLogFlops(2.0*pnz);CHKERRQ(ierr); 839 } 840 841 /* off-diagonal portion of A */ 842 anz = aoi[i+1] - aoi[i]; 843 aoj = ao->j + aoi[i]; 844 aoa = ao->a + aoi[i]; 845 for (j=0; j<anz; j++) { 846 row = aoj[j]; 847 pnz = pi_oth[row+1] - pi_oth[row]; 848 pj = pj_oth + pi_oth[row]; 849 pa = pa_oth + pi_oth[row]; 850 /* perform sparse axpy */ 851 valtmp = aoa[j]; 852 nextp = 0; 853 for (k=0; nextp<pnz; k++) { 854 if (apJ[k] == pj[nextp]) { /* column of AP == column of P */ 855 apa_sparse[k] += valtmp*pa[nextp++]; 856 } 857 } 858 ierr = PetscLogFlops(2.0*pnz);CHKERRQ(ierr); 859 } 860 861 /* set values in C */ 862 cdnz = cd->i[i+1] - cd->i[i]; 863 conz = co->i[i+1] - co->i[i]; 864 865 /* 1st off-diagoanl part of C */ 866 ca = coa + co->i[i]; 867 k = 0; 868 for (k0=0; k0<conz; k0++){ 869 if (apJ[k] >= cstart) break; 870 ca[k0] = apa_sparse[k]; 871 apa_sparse[k] = 0.0; 872 k++; 873 } 874 875 /* diagonal part of C */ 876 ca = cda + cd->i[i]; 877 for (k1=0; k1<cdnz; k1++){ 878 ca[k1] = apa_sparse[k]; 879 apa_sparse[k] = 0.0; 880 k++; 881 } 882 883 /* 2nd off-diagoanl part of C */ 884 ca = coa + co->i[i]; 885 for (; k0<conz; k0++){ 886 ca[k0] = apa_sparse[k]; 887 apa_sparse[k] = 0.0; 888 k++; 889 } 890 } 891 ierr = MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 892 ierr = MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 893 PetscFunctionReturn(0); 894 } 895 896 /* same as MatMatMultSymbolic_MPIAIJ_MPIAIJ(), except using LLCondensed to avoid O(BN) memory requirement */ 897 #undef __FUNCT__ 898 #define __FUNCT__ "MatMatMultSymbolic_MPIAIJ_MPIAIJ_Scalable" 899 PetscErrorCode MatMatMultSymbolic_MPIAIJ_MPIAIJ_Scalable(Mat A,Mat P,PetscReal fill,Mat *C) 900 { 901 PetscErrorCode ierr; 902 MPI_Comm comm=((PetscObject)A)->comm; 903 Mat Cmpi; 904 Mat_PtAPMPI *ptap; 905 PetscFreeSpaceList free_space=PETSC_NULL,current_space=PETSC_NULL; 906 Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data,*c; 907 Mat_SeqAIJ *ad=(Mat_SeqAIJ*)(a->A)->data,*ao=(Mat_SeqAIJ*)(a->B)->data,*p_loc,*p_oth; 908 PetscInt *pi_loc,*pj_loc,*pi_oth,*pj_oth,*dnz,*onz; 909 PetscInt *adi=ad->i,*adj=ad->j,*aoi=ao->i,*aoj=ao->j,rstart=A->rmap->rstart; 910 PetscInt i,pnz,row,*api,*apj,*Jptr,apnz,nspacedouble=0,j,nzi,*lnk,apnz_max=0; 911 PetscInt am=A->rmap->n,pN=P->cmap->N,pn=P->cmap->n,pm=P->rmap->n; 912 PetscInt nlnk_max,armax,prmax; 913 PetscReal afill; 914 PetscScalar *apa; 915 #if defined(DEBUG_MATMATMULT) 916 PetscMPIInt rank=a->rank; 917 #endif 918 919 PetscFunctionBegin; 920 #if defined(DEBUG_MATMATMULT) 921 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 922 if (!rank) ierr = PetscPrintf(PETSC_COMM_SELF,"[%d] call MatMatMultSymbolic_MPIAIJ_MPIAIJ_Scalable()...\n",rank); 923 #endif 924 925 /* create struct Mat_PtAPMPI and attached it to C later */ 926 ierr = PetscNew(Mat_PtAPMPI,&ptap);CHKERRQ(ierr); 927 928 /* get P_oth by taking rows of P (= non-zero cols of local A) from other processors */ 929 ierr = MatGetBrowsOfAoCols_MPIAIJ(A,P,MAT_INITIAL_MATRIX,&ptap->startsj,&ptap->startsj_r,&ptap->bufa,&ptap->P_oth);CHKERRQ(ierr); 930 #if defined(DEBUG_MATMATMULT) 931 if (!rank) ierr = PetscPrintf(PETSC_COMM_SELF,"[%d] P_oth is done...\n",rank); 932 #endif 933 /* get P_loc by taking all local rows of P */ 934 ierr = MatMPIAIJGetLocalMat(P,MAT_INITIAL_MATRIX,&ptap->P_loc);CHKERRQ(ierr); 935 936 p_loc = (Mat_SeqAIJ*)(ptap->P_loc)->data; 937 p_oth = (Mat_SeqAIJ*)(ptap->P_oth)->data; 938 pi_loc = p_loc->i; pj_loc = p_loc->j; 939 pi_oth = p_oth->i; pj_oth = p_oth->j; 940 941 #if defined(DEBUG_MATMATMULT) 942 if (!rank) ierr = PetscPrintf(PETSC_COMM_SELF,"[%d] P_loc is done, Annz %D * P_locnnz %D = %D...\n",rank,ad->i[am]+ao->i[am],p_loc->rmax,(ad->i[am]+ao->i[am])*p_loc->rmax); 943 #endif 944 945 /* first, compute symbolic AP = A_loc*P = A_diag*P_loc + A_off*P_oth */ 946 /*-------------------------------------------------------------------*/ 947 ierr = PetscMalloc((am+2)*sizeof(PetscInt),&api);CHKERRQ(ierr); 948 ptap->api = api; 949 api[0] = 0; 950 951 /* create and initialize a linked list */ 952 armax = ad->rmax+ao->rmax; 953 prmax = PetscMax(p_loc->rmax,p_oth->rmax); 954 nlnk_max = armax*prmax; 955 if (!nlnk_max || nlnk_max > pN) nlnk_max = pN; 956 #if defined(DEBUG_MATMATMULT) 957 if (!rank) ierr = PetscPrintf(PETSC_COMM_SELF,"[%d] pN %d; nlnk_max %d; Armax %d+%d=%d; Prmax Max(%d,%d)=%d\n",rank,pN,nlnk_max,ad->rmax,ao->rmax,armax,p_loc->rmax,p_oth->rmax,prmax); 958 #endif 959 ierr = PetscLLCondensedCreate_Scalable(nlnk_max,&lnk);CHKERRQ(ierr); 960 961 /* Initial FreeSpace size is fill*(nnz(A)+nnz(P)) */ 962 ierr = PetscFreeSpaceGet((PetscInt)(fill*(adi[am]+aoi[am]+pi_loc[pm])),&free_space);CHKERRQ(ierr); 963 current_space = free_space; 964 965 ierr = MatPreallocateInitialize(comm,am,pn,dnz,onz);CHKERRQ(ierr); 966 for (i=0; i<am; i++) { 967 apnz = 0; 968 /* diagonal portion of A */ 969 nzi = adi[i+1] - adi[i]; 970 for (j=0; j<nzi; j++){ 971 row = *adj++; 972 pnz = pi_loc[row+1] - pi_loc[row]; 973 Jptr = pj_loc + pi_loc[row]; 974 /* add non-zero cols of P into the sorted linked list lnk */ 975 ierr = PetscLLCondensedAddSorted_Scalable(pnz,Jptr,lnk);CHKERRQ(ierr); 976 } 977 /* off-diagonal portion of A */ 978 nzi = aoi[i+1] - aoi[i]; 979 for (j=0; j<nzi; j++){ 980 row = *aoj++; 981 pnz = pi_oth[row+1] - pi_oth[row]; 982 Jptr = pj_oth + pi_oth[row]; 983 ierr = PetscLLCondensedAddSorted_Scalable(pnz,Jptr,lnk);CHKERRQ(ierr); 984 } 985 986 apnz = *lnk; 987 api[i+1] = api[i] + apnz; 988 if (apnz > apnz_max) apnz_max = apnz; 989 990 /* if free space is not available, double the total space in the list */ 991 if (current_space->local_remaining<apnz) { 992 ierr = PetscFreeSpaceGet(apnz+current_space->total_array_size,¤t_space);CHKERRQ(ierr); 993 nspacedouble++; 994 } 995 996 /* Copy data into free space, then initialize lnk */ 997 ierr = PetscLLCondensedClean_Scalable(apnz,current_space->array,lnk);CHKERRQ(ierr); 998 ierr = MatPreallocateSet(i+rstart,apnz,current_space->array,dnz,onz);CHKERRQ(ierr); 999 current_space->array += apnz; 1000 current_space->local_used += apnz; 1001 current_space->local_remaining -= apnz; 1002 } 1003 1004 /* Allocate space for apj, initialize apj, and */ 1005 /* destroy list of free space and other temporary array(s) */ 1006 ierr = PetscMalloc((api[am]+1)*sizeof(PetscInt),&ptap->apj);CHKERRQ(ierr); 1007 apj = ptap->apj; 1008 ierr = PetscFreeSpaceContiguous(&free_space,ptap->apj);CHKERRQ(ierr); 1009 ierr = PetscLLCondensedDestroy_Scalable(lnk);CHKERRQ(ierr); 1010 #if defined(DEBUG_MATMATMULT) 1011 if (!rank) ierr = PetscPrintf(PETSC_COMM_SELF,"[%d] AP is done..., apnz_max %d\n",rank,apnz_max); 1012 #endif 1013 1014 /* create and assemble symbolic parallel matrix Cmpi */ 1015 /*----------------------------------------------------*/ 1016 ierr = MatCreate(comm,&Cmpi);CHKERRQ(ierr); 1017 ierr = MatSetSizes(Cmpi,am,pn,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 1018 ierr = MatSetType(Cmpi,MATMPIAIJ);CHKERRQ(ierr); 1019 ierr = MatMPIAIJSetPreallocation(Cmpi,0,dnz,0,onz);CHKERRQ(ierr); 1020 ierr = MatPreallocateFinalize(dnz,onz);CHKERRQ(ierr); 1021 ierr = MatSetBlockSize(Cmpi,1);CHKERRQ(ierr); 1022 1023 /* malloc apa for assembly Cmpi */ 1024 ierr = PetscMalloc(apnz_max*sizeof(PetscScalar),&apa);CHKERRQ(ierr); 1025 ierr = PetscMemzero(apa,apnz_max*sizeof(PetscScalar));CHKERRQ(ierr); 1026 ptap->apa = apa; 1027 for (i=0; i<am; i++){ 1028 row = i + rstart; 1029 apnz = api[i+1] - api[i]; 1030 ierr = MatSetValues(Cmpi,1,&row,apnz,apj,apa,INSERT_VALUES);CHKERRQ(ierr); 1031 apj += apnz; 1032 } 1033 ierr = MatAssemblyBegin(Cmpi,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1034 ierr = MatAssemblyEnd(Cmpi,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1035 1036 ptap->destroy = Cmpi->ops->destroy; 1037 ptap->duplicate = Cmpi->ops->duplicate; 1038 Cmpi->ops->matmultnumeric = MatMatMultNumeric_MPIAIJ_MPIAIJ_Scalable; 1039 Cmpi->ops->destroy = MatDestroy_MPIAIJ_MatMatMult; 1040 Cmpi->ops->duplicate = MatDuplicate_MPIAIJ_MatMatMult; 1041 1042 /* attach the supporting struct to Cmpi for reuse */ 1043 c = (Mat_MPIAIJ*)Cmpi->data; 1044 c->ptap = ptap; 1045 1046 *C = Cmpi; 1047 1048 /* set MatInfo */ 1049 afill = (PetscReal)api[am]/(adi[am]+aoi[am]+pi_loc[pm]) + 1.e-5; 1050 if (afill < 1.0) afill = 1.0; 1051 Cmpi->info.mallocs = nspacedouble; 1052 Cmpi->info.fill_ratio_given = fill; 1053 Cmpi->info.fill_ratio_needed = afill; 1054 1055 #if defined(PETSC_USE_INFO) 1056 if (api[am]) { 1057 ierr = PetscInfo3(Cmpi,"Reallocs %D; Fill ratio: given %G needed %G.\n",nspacedouble,fill,afill);CHKERRQ(ierr); 1058 ierr = PetscInfo1(Cmpi,"Use MatMatMult(A,B,MatReuse,%G,&C) for best performance.;\n",afill);CHKERRQ(ierr); 1059 } else { 1060 ierr = PetscInfo(Cmpi,"Empty matrix product\n");CHKERRQ(ierr); 1061 } 1062 #endif 1063 PetscFunctionReturn(0); 1064 } 1065 1066 /*-------------------------------------------------------------------------*/ 1067 #undef __FUNCT__ 1068 #define __FUNCT__ "MatTransposeMatMult_MPIAIJ_MPIAIJ" 1069 PetscErrorCode MatTransposeMatMult_MPIAIJ_MPIAIJ(Mat P,Mat A,MatReuse scall,PetscReal fill,Mat *C) 1070 { 1071 PetscErrorCode ierr; 1072 1073 PetscFunctionBegin; 1074 if (scall == MAT_INITIAL_MATRIX){ 1075 ierr = MatTransposeMatMultSymbolic_MPIAIJ_MPIAIJ(P,A,fill,C);CHKERRQ(ierr); 1076 } 1077 ierr = MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ(P,A,*C);CHKERRQ(ierr); 1078 PetscFunctionReturn(0); 1079 } 1080 1081 #undef __FUNCT__ 1082 #define __FUNCT__ "MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ" 1083 PetscErrorCode MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ(Mat P,Mat A,Mat C) 1084 { 1085 PetscErrorCode ierr; 1086 Mat_Merge_SeqsToMPI *merge; 1087 Mat_MPIAIJ *p=(Mat_MPIAIJ*)P->data,*c=(Mat_MPIAIJ*)C->data; 1088 Mat_SeqAIJ *pd=(Mat_SeqAIJ*)(p->A)->data,*po=(Mat_SeqAIJ*)(p->B)->data; 1089 Mat_PtAPMPI *ptap; 1090 PetscInt *adj,*apJ; 1091 PetscInt i,j,k,anz,pnz,row,*cj; 1092 MatScalar *ada,*apa,*ca,valtmp; 1093 PetscInt am=A->rmap->n,cm=C->rmap->n,pon=(p->B)->cmap->n; 1094 MPI_Comm comm=((PetscObject)C)->comm; 1095 PetscMPIInt size,rank,taga,*len_s; 1096 PetscInt *owners,proc,nrows,**buf_ri_k,**nextrow,**nextci; 1097 PetscInt **buf_ri,**buf_rj; 1098 PetscInt cnz=0,*bj_i,*bi,*bj,bnz,nextcj; /* bi,bj,ba: local array of C(mpi mat) */ 1099 MPI_Request *s_waits,*r_waits; 1100 MPI_Status *status; 1101 MatScalar **abuf_r,*ba_i,*pA,*coa,*ba; 1102 PetscInt *api,*apj,*coi,*coj; 1103 PetscInt *poJ=po->j,*pdJ=pd->j; 1104 PetscInt sparse_axpy; 1105 Mat A_loc; 1106 Mat_SeqAIJ *a_loc; 1107 1108 PetscFunctionBegin; 1109 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 1110 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 1111 ierr = MPI_Barrier(comm);CHKERRQ(ierr); 1112 1113 ptap = c->ptap; 1114 merge = ptap->merge; 1115 1116 /* 2) compute numeric C_seq = P_loc^T*A_loc*P - dominating part */ 1117 /*--------------------------------------------------------------*/ 1118 /* get data from symbolic products */ 1119 coi = merge->coi; coj = merge->coj; 1120 ierr = PetscMalloc((coi[pon]+1)*sizeof(MatScalar),&coa);CHKERRQ(ierr); 1121 ierr = PetscMemzero(coa,coi[pon]*sizeof(MatScalar));CHKERRQ(ierr); 1122 1123 bi = merge->bi; bj = merge->bj; 1124 owners = merge->rowmap->range; 1125 ierr = PetscMalloc((bi[cm]+1)*sizeof(MatScalar),&ba);CHKERRQ(ierr); 1126 ierr = PetscMemzero(ba,bi[cm]*sizeof(MatScalar));CHKERRQ(ierr); 1127 1128 /* get A_loc by taking all local rows of A */ 1129 A_loc = ptap->A_loc; 1130 ierr = MatMPIAIJGetLocalMat(A,MAT_REUSE_MATRIX,&A_loc);CHKERRQ(ierr); 1131 a_loc = (Mat_SeqAIJ*)(A_loc)->data; 1132 api = a_loc->i; 1133 apj = a_loc->j; 1134 1135 ierr = PetscMalloc((A->cmap->N)*sizeof(PetscScalar),&apa);CHKERRQ(ierr); 1136 ierr = PetscMemzero(apa,A->cmap->N*sizeof(PetscScalar));CHKERRQ(ierr); 1137 1138 for (i=0; i<am; i++) { 1139 /* 2-a) put A[i,:] to dense array apa */ 1140 anz = api[i+1] - api[i]; 1141 adj = apj + api[i]; 1142 ada = a_loc->a + api[i]; 1143 for (j=0; j<anz; j++){ 1144 apa[adj[j]] = ada[j]; 1145 } 1146 1147 /* 2-b) Compute Cseq = P_loc[i,:]^T*A[i,:] using outer product */ 1148 /*--------------------------------------------------------------*/ 1149 /* put the value into Co=(p->B)^T*A (off-diagonal part, send to others) */ 1150 pnz = po->i[i+1] - po->i[i]; 1151 poJ = po->j + po->i[i]; 1152 pA = po->a + po->i[i]; 1153 for (j=0; j<pnz; j++){ 1154 row = poJ[j]; 1155 cnz = coi[row+1] - coi[row]; 1156 cj = coj + coi[row]; 1157 ca = coa + coi[row]; 1158 /* perform dense axpy */ 1159 valtmp = pA[j]; 1160 for (k=0; k<cnz; k++) { 1161 ca[k] += valtmp*apa[cj[k]]; 1162 } 1163 ierr = PetscLogFlops(2.0*cnz);CHKERRQ(ierr); 1164 } 1165 1166 /* put the value into Cd (diagonal part) */ 1167 pnz = pd->i[i+1] - pd->i[i]; 1168 pdJ = pd->j + pd->i[i]; 1169 pA = pd->a + pd->i[i]; 1170 for (j=0; j<pnz; j++){ 1171 row = pdJ[j]; 1172 cnz = bi[row+1] - bi[row]; 1173 cj = bj + bi[row]; 1174 ca = ba + bi[row]; 1175 /* perform dense axpy */ 1176 valtmp = pA[j]; 1177 for (k=0; k<cnz; k++) { 1178 ca[k] += valtmp*apa[cj[k]]; 1179 } 1180 ierr = PetscLogFlops(2.0*cnz);CHKERRQ(ierr); 1181 } 1182 1183 /* zero the current row of A*P */ 1184 apJ = apj + api[i]; 1185 for (k=0; k<anz; k++) apa[apJ[k]] = 0.0; 1186 } 1187 1188 /* 3) send and recv matrix values coa */ 1189 /*------------------------------------*/ 1190 buf_ri = merge->buf_ri; 1191 buf_rj = merge->buf_rj; 1192 len_s = merge->len_s; 1193 ierr = PetscCommGetNewTag(comm,&taga);CHKERRQ(ierr); 1194 ierr = PetscPostIrecvScalar(comm,taga,merge->nrecv,merge->id_r,merge->len_r,&abuf_r,&r_waits);CHKERRQ(ierr); 1195 1196 ierr = PetscMalloc2(merge->nsend+1,MPI_Request,&s_waits,size,MPI_Status,&status);CHKERRQ(ierr); 1197 for (proc=0,k=0; proc<size; proc++){ 1198 if (!len_s[proc]) continue; 1199 i = merge->owners_co[proc]; 1200 ierr = MPI_Isend(coa+coi[i],len_s[proc],MPIU_MATSCALAR,proc,taga,comm,s_waits+k);CHKERRQ(ierr); 1201 k++; 1202 } 1203 if (merge->nrecv) {ierr = MPI_Waitall(merge->nrecv,r_waits,status);CHKERRQ(ierr);} 1204 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,s_waits,status);CHKERRQ(ierr);} 1205 1206 ierr = PetscFree2(s_waits,status);CHKERRQ(ierr); 1207 ierr = PetscFree(r_waits);CHKERRQ(ierr); 1208 ierr = PetscFree(coa);CHKERRQ(ierr); 1209 1210 /* 4) insert local Cseq and received values into Cmpi */ 1211 /*----------------------------------------------------*/ 1212 ierr = PetscMalloc3(merge->nrecv,PetscInt**,&buf_ri_k,merge->nrecv,PetscInt*,&nextrow,merge->nrecv,PetscInt*,&nextci);CHKERRQ(ierr); 1213 for (k=0; k<merge->nrecv; k++){ 1214 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure -- memory corruption ???*/ 1215 nrows = *(buf_ri_k[k]); 1216 nextrow[k] = buf_ri_k[k]+1; /* next row number of k-th recved i-structure */ 1217 nextci[k] = buf_ri_k[k] + (nrows + 1);/* poins to the next i-structure of k-th recved i-structure */ 1218 } 1219 1220 for (i=0; i<cm; i++) { 1221 row = owners[rank] + i; /* global row index of C_seq */ 1222 bj_i = bj + bi[i]; /* col indices of the i-th row of C */ 1223 ba_i = ba + bi[i]; 1224 bnz = bi[i+1] - bi[i]; 1225 /* add received vals into ba */ 1226 for (k=0; k<merge->nrecv; k++){ /* k-th received message */ 1227 /* i-th row */ 1228 if (i == *nextrow[k]) { 1229 cnz = *(nextci[k]+1) - *nextci[k]; 1230 cj = buf_rj[k] + *(nextci[k]); 1231 ca = abuf_r[k] + *(nextci[k]); 1232 nextcj = 0; 1233 for (j=0; nextcj<cnz; j++){ 1234 if (bj_i[j] == cj[nextcj]){ /* bcol == ccol */ 1235 ba_i[j] += ca[nextcj++]; 1236 } 1237 } 1238 nextrow[k]++; nextci[k]++; 1239 ierr = PetscLogFlops(2.0*cnz);CHKERRQ(ierr); 1240 } 1241 } 1242 ierr = MatSetValues(C,1,&row,bnz,bj_i,ba_i,INSERT_VALUES);CHKERRQ(ierr); 1243 } 1244 ierr = MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1245 ierr = MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1246 1247 ierr = PetscFree(ba);CHKERRQ(ierr); 1248 ierr = PetscFree(abuf_r[0]);CHKERRQ(ierr); 1249 ierr = PetscFree(abuf_r);CHKERRQ(ierr); 1250 ierr = PetscFree3(buf_ri_k,nextrow,nextci);CHKERRQ(ierr); 1251 ierr = PetscFree(apa);CHKERRQ(ierr); 1252 PetscFunctionReturn(0); 1253 } 1254 1255 /* This routine is modified from MatPtAPSymbolic_MPIAIJ_MPIAIJ() */ 1256 #undef __FUNCT__ 1257 #define __FUNCT__ "MatTransposeMatMultSymbolic_MPIAIJ_MPIAIJ" 1258 PetscErrorCode MatTransposeMatMultSymbolic_MPIAIJ_MPIAIJ(Mat P,Mat A,PetscReal fill,Mat *C) 1259 { 1260 PetscErrorCode ierr; 1261 Mat Cmpi; 1262 Mat_PtAPMPI *ptap; 1263 PetscFreeSpaceList free_space=PETSC_NULL,current_space=PETSC_NULL; 1264 Mat_MPIAIJ *p=(Mat_MPIAIJ*)P->data,*c; 1265 PetscInt *pdti,*pdtj,*poti,*potj,*ptJ; 1266 PetscInt nnz; 1267 PetscInt nlnk,*lnk,*owners_co,*coi,*coj,i,k,pnz,row; 1268 PetscInt am=A->rmap->n,pn=P->cmap->n; 1269 PetscBT lnkbt; 1270 MPI_Comm comm=((PetscObject)A)->comm; 1271 PetscMPIInt size,rank,tagi,tagj,*len_si,*len_s,*len_ri; 1272 PetscInt **buf_rj,**buf_ri,**buf_ri_k; 1273 PetscInt len,proc,*dnz,*onz,*owners; 1274 PetscInt nzi,*bi,*bj; 1275 PetscInt nrows,*buf_s,*buf_si,*buf_si_i,**nextrow,**nextci; 1276 MPI_Request *swaits,*rwaits; 1277 MPI_Status *sstatus,rstatus; 1278 Mat_Merge_SeqsToMPI *merge; 1279 PetscInt *api,*apj,*Jptr,apnz,*prmap=p->garray,pon,nspacedouble=0,j,ap_rmax=0; 1280 PetscReal afill=1.0,afill_tmp; 1281 PetscInt rstart = P->cmap->rstart,rmax,aN=A->cmap->N; 1282 PetscScalar *vals; 1283 Mat A_loc; 1284 Mat_SeqAIJ *a_loc; 1285 1286 PetscFunctionBegin; 1287 /* check if matrix local sizes are compatible */ 1288 if (A->rmap->rstart != P->rmap->rstart || A->rmap->rend != P->rmap->rend){ 1289 SETERRQ4(comm,PETSC_ERR_ARG_SIZ,"Matrix local dimensions are incompatible, A (%D, %D) != P (%D,%D)",A->rmap->rstart,A->rmap->rend,P->rmap->rstart,P->rmap->rend); 1290 } 1291 1292 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 1293 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 1294 #if defined(DEBUG_MATTrMatMult) 1295 ierr = PetscSynchronizedPrintf(comm,"[%d] TransposeMatMultSymbolic P: %d %d, %d %d; A %d %d, %d %d\n",rank,P->rmap->N,pN,P->rmap->n,P->cmap->n,A->rmap->N,aN,A->rmap->n,A->cmap->n); 1296 ierr = PetscSynchronizedFlush(comm);CHKERRQ(ierr); 1297 #endif 1298 1299 /* create struct Mat_PtAPMPI and attached it to C later */ 1300 ierr = PetscNew(Mat_PtAPMPI,&ptap);CHKERRQ(ierr); 1301 1302 /* get A_loc by taking all local rows of A */ 1303 ierr = MatMPIAIJGetLocalMat(A,MAT_INITIAL_MATRIX,&A_loc);CHKERRQ(ierr); 1304 ptap->A_loc = A_loc; 1305 a_loc = (Mat_SeqAIJ*)(A_loc)->data; 1306 api = a_loc->i; 1307 apj = a_loc->j; 1308 1309 /* determine symbolic Co=(p->B)^T*A - send to others */ 1310 /*----------------------------------------------------*/ 1311 ierr = MatGetSymbolicTranspose_SeqAIJ(p->B,&poti,&potj);CHKERRQ(ierr); 1312 1313 /* then, compute symbolic Co = (p->B)^T*A */ 1314 pon = (p->B)->cmap->n; /* total num of rows to be sent to other processors 1315 >= (num of nonzero rows of C_seq) - pn */ 1316 ierr = PetscMalloc((pon+1)*sizeof(PetscInt),&coi);CHKERRQ(ierr); 1317 coi[0] = 0; 1318 1319 /* set initial free space to be fill*(nnz(p->B) + nnz(A)) */ 1320 nnz = fill*(poti[pon] + api[am]); 1321 ierr = PetscFreeSpaceGet(nnz,&free_space); 1322 current_space = free_space; 1323 #if defined(DEBUG_MATTrMatMult) 1324 ierr = PetscSynchronizedPrintf(comm, "[%d] nnz = fill %g *(%d + %d)\n",rank,fill,poti[pon],api[am]);CHKERRQ(ierr); 1325 ierr = PetscSynchronizedFlush(comm);CHKERRQ(ierr); 1326 #endif 1327 /* create and initialize a linked list */ 1328 nlnk = aN+1; 1329 ierr = PetscLLCreate(aN,aN,nlnk,lnk,lnkbt);CHKERRQ(ierr); 1330 1331 for (i=0; i<pon; i++) { 1332 nnz = 0; 1333 pnz = poti[i+1] - poti[i]; 1334 ptJ = potj + poti[i]; 1335 for (j=0; j<pnz; j++){ 1336 row = ptJ[j]; /* row of A_loc == col of Pot */ 1337 apnz = api[row+1] - api[row]; 1338 Jptr = apj + api[row]; 1339 /* add non-zero cols of AP into the sorted linked list lnk */ 1340 ierr = PetscLLAddSorted(apnz,Jptr,aN,nlnk,lnk,lnkbt);CHKERRQ(ierr); 1341 nnz += nlnk; 1342 } 1343 1344 /* If free space is not available, double the total space in the list */ 1345 if (current_space->local_remaining<nnz) { 1346 ierr = PetscFreeSpaceGet(nnz+current_space->total_array_size,¤t_space);CHKERRQ(ierr); 1347 nspacedouble++; 1348 } 1349 1350 /* Copy data into free space, and zero out denserows */ 1351 ierr = PetscLLClean(aN,aN,nnz,lnk,current_space->array,lnkbt);CHKERRQ(ierr); 1352 current_space->array += nnz; 1353 current_space->local_used += nnz; 1354 current_space->local_remaining -= nnz; 1355 coi[i+1] = coi[i] + nnz; 1356 } 1357 1358 ierr = PetscMalloc((coi[pon]+1)*sizeof(PetscInt),&coj);CHKERRQ(ierr); 1359 ierr = PetscFreeSpaceContiguous(&free_space,coj);CHKERRQ(ierr); 1360 afill_tmp = (PetscReal)coi[pon]/(poti[pon] + api[am]); 1361 if (afill_tmp > afill) afill = afill_tmp; 1362 1363 /* send j-array (coj) of Co to other processors */ 1364 /*----------------------------------------------*/ 1365 /* determine row ownership */ 1366 ierr = PetscNew(Mat_Merge_SeqsToMPI,&merge);CHKERRQ(ierr); 1367 ierr = PetscLayoutCreate(comm,&merge->rowmap);CHKERRQ(ierr); 1368 merge->rowmap->n = pn; 1369 merge->rowmap->bs = 1; 1370 ierr = PetscLayoutSetUp(merge->rowmap);CHKERRQ(ierr); 1371 owners = merge->rowmap->range; 1372 1373 /* determine the number of messages to send, their lengths */ 1374 ierr = PetscMalloc(size*sizeof(PetscMPIInt),&len_si);CHKERRQ(ierr); 1375 ierr = PetscMemzero(len_si,size*sizeof(PetscMPIInt));CHKERRQ(ierr); 1376 ierr = PetscMalloc(size*sizeof(PetscMPIInt),&merge->len_s);CHKERRQ(ierr); 1377 len_s = merge->len_s; 1378 merge->nsend = 0; 1379 1380 ierr = PetscMalloc((size+2)*sizeof(PetscInt),&owners_co);CHKERRQ(ierr); 1381 ierr = PetscMemzero(len_s,size*sizeof(PetscMPIInt));CHKERRQ(ierr); 1382 1383 proc = 0; 1384 for (i=0; i<pon; i++){ 1385 while (prmap[i] >= owners[proc+1]) proc++; 1386 len_si[proc]++; /* num of rows in Co to be sent to [proc] */ 1387 len_s[proc] += coi[i+1] - coi[i]; 1388 } 1389 1390 len = 0; /* max length of buf_si[] */ 1391 owners_co[0] = 0; 1392 for (proc=0; proc<size; proc++){ 1393 owners_co[proc+1] = owners_co[proc] + len_si[proc]; 1394 if (len_si[proc]){ 1395 merge->nsend++; 1396 len_si[proc] = 2*(len_si[proc] + 1); 1397 len += len_si[proc]; 1398 } 1399 } 1400 1401 /* determine the number and length of messages to receive for coi and coj */ 1402 ierr = PetscGatherNumberOfMessages(comm,PETSC_NULL,len_s,&merge->nrecv);CHKERRQ(ierr); 1403 ierr = PetscGatherMessageLengths2(comm,merge->nsend,merge->nrecv,len_s,len_si,&merge->id_r,&merge->len_r,&len_ri);CHKERRQ(ierr); 1404 1405 /* post the Irecv and Isend of coj */ 1406 ierr = PetscCommGetNewTag(comm,&tagj);CHKERRQ(ierr); 1407 ierr = PetscPostIrecvInt(comm,tagj,merge->nrecv,merge->id_r,merge->len_r,&buf_rj,&rwaits);CHKERRQ(ierr); 1408 ierr = PetscMalloc((merge->nsend+1)*sizeof(MPI_Request),&swaits);CHKERRQ(ierr); 1409 for (proc=0, k=0; proc<size; proc++){ 1410 if (!len_s[proc]) continue; 1411 i = owners_co[proc]; 1412 ierr = MPI_Isend(coj+coi[i],len_s[proc],MPIU_INT,proc,tagj,comm,swaits+k);CHKERRQ(ierr); 1413 k++; 1414 } 1415 1416 /* receives and sends of coj are complete */ 1417 ierr = PetscMalloc(size*sizeof(MPI_Status),&sstatus);CHKERRQ(ierr); 1418 for (i=0; i<merge->nrecv; i++){ 1419 ierr = MPI_Waitany(merge->nrecv,rwaits,&j,&rstatus);CHKERRQ(ierr); 1420 } 1421 ierr = PetscFree(rwaits);CHKERRQ(ierr); 1422 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,swaits,sstatus);CHKERRQ(ierr);} 1423 1424 /* send and recv coi */ 1425 /*-------------------*/ 1426 ierr = PetscCommGetNewTag(comm,&tagi);CHKERRQ(ierr); 1427 ierr = PetscPostIrecvInt(comm,tagi,merge->nrecv,merge->id_r,len_ri,&buf_ri,&rwaits);CHKERRQ(ierr); 1428 ierr = PetscMalloc((len+1)*sizeof(PetscInt),&buf_s);CHKERRQ(ierr); 1429 buf_si = buf_s; /* points to the beginning of k-th msg to be sent */ 1430 for (proc=0,k=0; proc<size; proc++){ 1431 if (!len_s[proc]) continue; 1432 /* form outgoing message for i-structure: 1433 buf_si[0]: nrows to be sent 1434 [1:nrows]: row index (global) 1435 [nrows+1:2*nrows+1]: i-structure index 1436 */ 1437 /*-------------------------------------------*/ 1438 nrows = len_si[proc]/2 - 1; 1439 buf_si_i = buf_si + nrows+1; 1440 buf_si[0] = nrows; 1441 buf_si_i[0] = 0; 1442 nrows = 0; 1443 for (i=owners_co[proc]; i<owners_co[proc+1]; i++){ 1444 nzi = coi[i+1] - coi[i]; 1445 buf_si_i[nrows+1] = buf_si_i[nrows] + nzi; /* i-structure */ 1446 buf_si[nrows+1] =prmap[i] -owners[proc]; /* local row index */ 1447 nrows++; 1448 } 1449 ierr = MPI_Isend(buf_si,len_si[proc],MPIU_INT,proc,tagi,comm,swaits+k);CHKERRQ(ierr); 1450 k++; 1451 buf_si += len_si[proc]; 1452 } 1453 i = merge->nrecv; 1454 while (i--) { 1455 ierr = MPI_Waitany(merge->nrecv,rwaits,&j,&rstatus);CHKERRQ(ierr); 1456 } 1457 ierr = PetscFree(rwaits);CHKERRQ(ierr); 1458 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,swaits,sstatus);CHKERRQ(ierr);} 1459 ierr = PetscFree(len_si);CHKERRQ(ierr); 1460 ierr = PetscFree(len_ri);CHKERRQ(ierr); 1461 ierr = PetscFree(swaits);CHKERRQ(ierr); 1462 ierr = PetscFree(sstatus);CHKERRQ(ierr); 1463 ierr = PetscFree(buf_s);CHKERRQ(ierr); 1464 1465 /* compute the local portion of C (mpi mat) */ 1466 /*------------------------------------------*/ 1467 ierr = MatGetSymbolicTranspose_SeqAIJ(p->A,&pdti,&pdtj);CHKERRQ(ierr); 1468 1469 /* allocate bi array and free space for accumulating nonzero column info */ 1470 ierr = PetscMalloc((pn+1)*sizeof(PetscInt),&bi);CHKERRQ(ierr); 1471 bi[0] = 0; 1472 1473 /* set initial free space to be fill*(nnz(P) + nnz(AP)) */ 1474 nnz = fill*(pdti[pn] + poti[pon] + api[am]); 1475 ierr = PetscFreeSpaceGet(nnz,&free_space); 1476 current_space = free_space; 1477 1478 #if defined(DEBUG_MATTrMatMult) 1479 ierr = PetscPrintf(PETSC_COMM_SELF,"[%d] nnz=%d + %d + %d\n",rank,pdti[pn],poti[pon],api[am]); 1480 #endif 1481 1482 ierr = PetscMalloc3(merge->nrecv,PetscInt**,&buf_ri_k,merge->nrecv,PetscInt*,&nextrow,merge->nrecv,PetscInt*,&nextci);CHKERRQ(ierr); 1483 for (k=0; k<merge->nrecv; k++){ 1484 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 1485 nrows = *buf_ri_k[k]; 1486 nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */ 1487 nextci[k] = buf_ri_k[k] + (nrows + 1);/* poins to the next i-structure of k-th recved i-structure */ 1488 } 1489 1490 ierr = MatPreallocateInitialize(comm,pn,A->cmap->n,dnz,onz);CHKERRQ(ierr); 1491 rmax = 0; 1492 for (i=0; i<pn; i++) { 1493 /* add pdt[i,:]*AP into lnk */ 1494 nnz = 0; 1495 pnz = pdti[i+1] - pdti[i]; 1496 ptJ = pdtj + pdti[i]; 1497 for (j=0; j<pnz; j++){ 1498 row = ptJ[j]; /* row of AP == col of Pt */ 1499 apnz = api[row+1] - api[row]; 1500 Jptr = apj + api[row]; 1501 /* add non-zero cols of AP into the sorted linked list lnk */ 1502 ierr = PetscLLAddSorted(apnz,Jptr,aN,nlnk,lnk,lnkbt);CHKERRQ(ierr); 1503 nnz += nlnk; 1504 } 1505 /* add received col data into lnk */ 1506 for (k=0; k<merge->nrecv; k++){ /* k-th received message */ 1507 if (i == *nextrow[k]) { /* i-th row */ 1508 nzi = *(nextci[k]+1) - *nextci[k]; 1509 Jptr = buf_rj[k] + *nextci[k]; 1510 ierr = PetscLLAddSorted(nzi,Jptr,aN,nlnk,lnk,lnkbt);CHKERRQ(ierr); 1511 nnz += nlnk; 1512 nextrow[k]++; nextci[k]++; 1513 } 1514 } 1515 1516 /* if free space is not available, make more free space */ 1517 if (current_space->local_remaining<nnz) { 1518 ierr = PetscFreeSpaceGet(nnz+current_space->total_array_size,¤t_space);CHKERRQ(ierr); 1519 nspacedouble++; 1520 } 1521 /* copy data into free space, then initialize lnk */ 1522 ierr = PetscLLClean(aN,aN,nnz,lnk,current_space->array,lnkbt);CHKERRQ(ierr); 1523 ierr = MatPreallocateSet(i+owners[rank],nnz,current_space->array,dnz,onz);CHKERRQ(ierr); 1524 current_space->array += nnz; 1525 current_space->local_used += nnz; 1526 current_space->local_remaining -= nnz; 1527 bi[i+1] = bi[i] + nnz; 1528 if (nnz > rmax) rmax = nnz; 1529 } 1530 ierr = PetscFree3(buf_ri_k,nextrow,nextci);CHKERRQ(ierr); 1531 1532 ierr = PetscMalloc((bi[pn]+1)*sizeof(PetscInt),&bj);CHKERRQ(ierr); 1533 ierr = PetscFreeSpaceContiguous(&free_space,bj);CHKERRQ(ierr); 1534 afill_tmp = (PetscReal)bi[pn]/(pdti[pn] + poti[pon] + api[am]); 1535 if (afill_tmp > afill) afill = afill_tmp; 1536 ierr = PetscLLDestroy(lnk,lnkbt);CHKERRQ(ierr); 1537 ierr = MatRestoreSymbolicTranspose_SeqAIJ(p->A,&pdti,&pdtj);CHKERRQ(ierr); 1538 ierr = MatRestoreSymbolicTranspose_SeqAIJ(p->B,&poti,&potj);CHKERRQ(ierr); 1539 1540 /* create symbolic parallel matrix Cmpi - why cannot be assembled in Numeric part */ 1541 /*----------------------------------------------------------------------------------*/ 1542 ierr = PetscMalloc((rmax+1)*sizeof(PetscScalar),&vals);CHKERRQ(ierr); 1543 ierr = PetscMemzero(vals,rmax*sizeof(PetscScalar));CHKERRQ(ierr); 1544 1545 ierr = MatCreate(comm,&Cmpi);CHKERRQ(ierr); 1546 ierr = MatSetSizes(Cmpi,pn,A->cmap->n,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 1547 ierr = MatSetType(Cmpi,MATMPIAIJ);CHKERRQ(ierr); 1548 ierr = MatMPIAIJSetPreallocation(Cmpi,0,dnz,0,onz);CHKERRQ(ierr); 1549 ierr = MatPreallocateFinalize(dnz,onz);CHKERRQ(ierr); 1550 ierr = MatSetBlockSize(Cmpi,1);CHKERRQ(ierr); 1551 for (i=0; i<pn; i++){ 1552 row = i + rstart; 1553 nnz = bi[i+1] - bi[i]; 1554 Jptr = bj + bi[i]; 1555 ierr = MatSetValues(Cmpi,1,&row,nnz,Jptr,vals,INSERT_VALUES);CHKERRQ(ierr); 1556 } 1557 ierr = MatAssemblyBegin(Cmpi,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1558 ierr = MatAssemblyEnd(Cmpi,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1559 ierr = PetscFree(vals);CHKERRQ(ierr); 1560 1561 merge->bi = bi; 1562 merge->bj = bj; 1563 merge->coi = coi; 1564 merge->coj = coj; 1565 merge->buf_ri = buf_ri; 1566 merge->buf_rj = buf_rj; 1567 merge->owners_co = owners_co; 1568 merge->destroy = Cmpi->ops->destroy; 1569 merge->duplicate = Cmpi->ops->duplicate; 1570 1571 Cmpi->ops->destroy = MatDestroy_MPIAIJ_PtAP; 1572 1573 /* attach the supporting struct to Cmpi for reuse */ 1574 c = (Mat_MPIAIJ*)Cmpi->data; 1575 c->ptap = ptap; 1576 ptap->api = PETSC_NULL; 1577 ptap->apj = PETSC_NULL; 1578 ptap->merge = merge; 1579 ptap->apnz_max = ap_rmax; 1580 1581 *C = Cmpi; 1582 #if defined(PETSC_USE_INFO) 1583 if (bi[pn] != 0) { 1584 ierr = PetscInfo3(Cmpi,"Reallocs %D; Fill ratio: given %G needed %G.\n",nspacedouble,fill,afill);CHKERRQ(ierr); 1585 ierr = PetscInfo1(Cmpi,"Use MatTransposeMatMult(A,B,MatReuse,%G,&C) for best performance.\n",afill);CHKERRQ(ierr); 1586 } else { 1587 ierr = PetscInfo(Cmpi,"Empty matrix product\n");CHKERRQ(ierr); 1588 } 1589 #endif 1590 PetscFunctionReturn(0); 1591 } 1592