1 /* 2 Defines matrix-matrix product routines for pairs of MPIAIJ matrices 3 C = A * B 4 */ 5 #include <../src/mat/impls/aij/seq/aij.h> /*I "petscmat.h" I*/ 6 #include <../src/mat/utils/freespace.h> 7 #include <../src/mat/impls/aij/mpi/mpiaij.h> 8 #include <petscbt.h> 9 #include <../src/mat/impls/dense/mpi/mpidense.h> 10 #include <petsc/private/vecimpl.h> 11 #include <petsc/private/sfimpl.h> 12 13 #if defined(PETSC_HAVE_HYPRE) 14 PETSC_INTERN PetscErrorCode MatMatMultSymbolic_AIJ_AIJ_wHYPRE(Mat, Mat, PetscReal, Mat); 15 #endif 16 17 PETSC_INTERN PetscErrorCode MatProductSymbolic_ABt_MPIAIJ_MPIAIJ(Mat C) 18 { 19 Mat_Product *product = C->product; 20 Mat B = product->B; 21 22 PetscFunctionBegin; 23 PetscCall(MatTranspose(B, MAT_INITIAL_MATRIX, &product->B)); 24 PetscCall(MatDestroy(&B)); 25 PetscCall(MatProductSymbolic_AB_MPIAIJ_MPIAIJ(C)); 26 PetscFunctionReturn(PETSC_SUCCESS); 27 } 28 29 PETSC_INTERN PetscErrorCode MatProductSymbolic_AB_MPIAIJ_MPIAIJ(Mat C) 30 { 31 Mat_Product *product = C->product; 32 Mat A = product->A, B = product->B; 33 MatProductAlgorithm alg = product->alg; 34 PetscReal fill = product->fill; 35 PetscBool flg; 36 37 PetscFunctionBegin; 38 /* scalable */ 39 PetscCall(PetscStrcmp(alg, "scalable", &flg)); 40 if (flg) { 41 PetscCall(MatMatMultSymbolic_MPIAIJ_MPIAIJ(A, B, fill, C)); 42 PetscFunctionReturn(PETSC_SUCCESS); 43 } 44 45 /* nonscalable */ 46 PetscCall(PetscStrcmp(alg, "nonscalable", &flg)); 47 if (flg) { 48 PetscCall(MatMatMultSymbolic_MPIAIJ_MPIAIJ_nonscalable(A, B, fill, C)); 49 PetscFunctionReturn(PETSC_SUCCESS); 50 } 51 52 /* seqmpi */ 53 PetscCall(PetscStrcmp(alg, "seqmpi", &flg)); 54 if (flg) { 55 PetscCall(MatMatMultSymbolic_MPIAIJ_MPIAIJ_seqMPI(A, B, fill, C)); 56 PetscFunctionReturn(PETSC_SUCCESS); 57 } 58 59 /* backend general code */ 60 PetscCall(PetscStrcmp(alg, "backend", &flg)); 61 if (flg) { 62 PetscCall(MatProductSymbolic_MPIAIJBACKEND(C)); 63 PetscFunctionReturn(PETSC_SUCCESS); 64 } 65 66 #if defined(PETSC_HAVE_HYPRE) 67 PetscCall(PetscStrcmp(alg, "hypre", &flg)); 68 if (flg) { 69 PetscCall(MatMatMultSymbolic_AIJ_AIJ_wHYPRE(A, B, fill, C)); 70 PetscFunctionReturn(PETSC_SUCCESS); 71 } 72 #endif 73 SETERRQ(PetscObjectComm((PetscObject)C), PETSC_ERR_SUP, "Mat Product Algorithm is not supported"); 74 } 75 76 PetscErrorCode MatDestroy_MPIAIJ_MatMatMult(void *data) 77 { 78 Mat_APMPI *ptap = (Mat_APMPI *)data; 79 80 PetscFunctionBegin; 81 PetscCall(PetscFree2(ptap->startsj_s, ptap->startsj_r)); 82 PetscCall(PetscFree(ptap->bufa)); 83 PetscCall(MatDestroy(&ptap->P_loc)); 84 PetscCall(MatDestroy(&ptap->P_oth)); 85 PetscCall(MatDestroy(&ptap->Pt)); 86 PetscCall(PetscFree(ptap->api)); 87 PetscCall(PetscFree(ptap->apj)); 88 PetscCall(PetscFree(ptap->apa)); 89 PetscCall(PetscFree(ptap)); 90 PetscFunctionReturn(PETSC_SUCCESS); 91 } 92 93 PetscErrorCode MatMatMultNumeric_MPIAIJ_MPIAIJ_nonscalable(Mat A, Mat P, Mat C) 94 { 95 Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data, *c = (Mat_MPIAIJ *)C->data; 96 Mat_SeqAIJ *ad = (Mat_SeqAIJ *)(a->A)->data, *ao = (Mat_SeqAIJ *)(a->B)->data; 97 Mat_SeqAIJ *cd = (Mat_SeqAIJ *)(c->A)->data, *co = (Mat_SeqAIJ *)(c->B)->data; 98 PetscScalar *cda = cd->a, *coa = co->a; 99 Mat_SeqAIJ *p_loc, *p_oth; 100 PetscScalar *apa, *ca; 101 PetscInt cm = C->rmap->n; 102 Mat_APMPI *ptap; 103 PetscInt *api, *apj, *apJ, i, k; 104 PetscInt cstart = C->cmap->rstart; 105 PetscInt cdnz, conz, k0, k1; 106 const PetscScalar *dummy; 107 MPI_Comm comm; 108 PetscMPIInt size; 109 110 PetscFunctionBegin; 111 MatCheckProduct(C, 3); 112 ptap = (Mat_APMPI *)C->product->data; 113 PetscCheck(ptap, PetscObjectComm((PetscObject)C), PETSC_ERR_ARG_WRONGSTATE, "PtAP cannot be computed. Missing data"); 114 PetscCall(PetscObjectGetComm((PetscObject)A, &comm)); 115 PetscCallMPI(MPI_Comm_size(comm, &size)); 116 PetscCheck(ptap->P_oth || size <= 1, PetscObjectComm((PetscObject)C), PETSC_ERR_ARG_WRONGSTATE, "AP cannot be reused. Do not call MatProductClear()"); 117 118 /* flag CPU mask for C */ 119 #if defined(PETSC_HAVE_DEVICE) 120 if (C->offloadmask != PETSC_OFFLOAD_UNALLOCATED) C->offloadmask = PETSC_OFFLOAD_CPU; 121 if (c->A->offloadmask != PETSC_OFFLOAD_UNALLOCATED) c->A->offloadmask = PETSC_OFFLOAD_CPU; 122 if (c->B->offloadmask != PETSC_OFFLOAD_UNALLOCATED) c->B->offloadmask = PETSC_OFFLOAD_CPU; 123 #endif 124 125 /* 1) get P_oth = ptap->P_oth and P_loc = ptap->P_loc */ 126 /* update numerical values of P_oth and P_loc */ 127 PetscCall(MatGetBrowsOfAoCols_MPIAIJ(A, P, MAT_REUSE_MATRIX, &ptap->startsj_s, &ptap->startsj_r, &ptap->bufa, &ptap->P_oth)); 128 PetscCall(MatMPIAIJGetLocalMat(P, MAT_REUSE_MATRIX, &ptap->P_loc)); 129 130 /* 2) compute numeric C_loc = A_loc*P = Ad*P_loc + Ao*P_oth */ 131 /* get data from symbolic products */ 132 p_loc = (Mat_SeqAIJ *)(ptap->P_loc)->data; 133 p_oth = NULL; 134 if (size > 1) p_oth = (Mat_SeqAIJ *)(ptap->P_oth)->data; 135 136 /* get apa for storing dense row A[i,:]*P */ 137 apa = ptap->apa; 138 139 api = ptap->api; 140 apj = ptap->apj; 141 /* trigger copy to CPU */ 142 PetscCall(MatSeqAIJGetArrayRead(a->A, &dummy)); 143 PetscCall(MatSeqAIJRestoreArrayRead(a->A, &dummy)); 144 PetscCall(MatSeqAIJGetArrayRead(a->B, &dummy)); 145 PetscCall(MatSeqAIJRestoreArrayRead(a->B, &dummy)); 146 for (i = 0; i < cm; i++) { 147 /* compute apa = A[i,:]*P */ 148 AProw_nonscalable(i, ad, ao, p_loc, p_oth, apa); 149 150 /* set values in C */ 151 apJ = apj + api[i]; 152 cdnz = cd->i[i + 1] - cd->i[i]; 153 conz = co->i[i + 1] - co->i[i]; 154 155 /* 1st off-diagonal part of C */ 156 ca = coa + co->i[i]; 157 k = 0; 158 for (k0 = 0; k0 < conz; k0++) { 159 if (apJ[k] >= cstart) break; 160 ca[k0] = apa[apJ[k]]; 161 apa[apJ[k++]] = 0.0; 162 } 163 164 /* diagonal part of C */ 165 ca = cda + cd->i[i]; 166 for (k1 = 0; k1 < cdnz; k1++) { 167 ca[k1] = apa[apJ[k]]; 168 apa[apJ[k++]] = 0.0; 169 } 170 171 /* 2nd off-diagonal part of C */ 172 ca = coa + co->i[i]; 173 for (; k0 < conz; k0++) { 174 ca[k0] = apa[apJ[k]]; 175 apa[apJ[k++]] = 0.0; 176 } 177 } 178 PetscCall(MatAssemblyBegin(C, MAT_FINAL_ASSEMBLY)); 179 PetscCall(MatAssemblyEnd(C, MAT_FINAL_ASSEMBLY)); 180 PetscFunctionReturn(PETSC_SUCCESS); 181 } 182 183 PetscErrorCode MatMatMultSymbolic_MPIAIJ_MPIAIJ_nonscalable(Mat A, Mat P, PetscReal fill, Mat C) 184 { 185 MPI_Comm comm; 186 PetscMPIInt size; 187 Mat_APMPI *ptap; 188 PetscFreeSpaceList free_space = NULL, current_space = NULL; 189 Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data; 190 Mat_SeqAIJ *ad = (Mat_SeqAIJ *)(a->A)->data, *ao = (Mat_SeqAIJ *)(a->B)->data, *p_loc, *p_oth; 191 PetscInt *pi_loc, *pj_loc, *pi_oth, *pj_oth, *dnz, *onz; 192 PetscInt *adi = ad->i, *adj = ad->j, *aoi = ao->i, *aoj = ao->j, rstart = A->rmap->rstart; 193 PetscInt *lnk, i, pnz, row, *api, *apj, *Jptr, apnz, nspacedouble = 0, j, nzi; 194 PetscInt am = A->rmap->n, pN = P->cmap->N, pn = P->cmap->n, pm = P->rmap->n; 195 PetscBT lnkbt; 196 PetscReal afill; 197 MatType mtype; 198 199 PetscFunctionBegin; 200 MatCheckProduct(C, 4); 201 PetscCheck(!C->product->data, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Extra product struct not empty"); 202 PetscCall(PetscObjectGetComm((PetscObject)A, &comm)); 203 PetscCallMPI(MPI_Comm_size(comm, &size)); 204 205 /* create struct Mat_APMPI and attached it to C later */ 206 PetscCall(PetscNew(&ptap)); 207 208 /* get P_oth by taking rows of P (= non-zero cols of local A) from other processors */ 209 PetscCall(MatGetBrowsOfAoCols_MPIAIJ(A, P, MAT_INITIAL_MATRIX, &ptap->startsj_s, &ptap->startsj_r, &ptap->bufa, &ptap->P_oth)); 210 211 /* get P_loc by taking all local rows of P */ 212 PetscCall(MatMPIAIJGetLocalMat(P, MAT_INITIAL_MATRIX, &ptap->P_loc)); 213 214 p_loc = (Mat_SeqAIJ *)(ptap->P_loc)->data; 215 pi_loc = p_loc->i; 216 pj_loc = p_loc->j; 217 if (size > 1) { 218 p_oth = (Mat_SeqAIJ *)(ptap->P_oth)->data; 219 pi_oth = p_oth->i; 220 pj_oth = p_oth->j; 221 } else { 222 p_oth = NULL; 223 pi_oth = NULL; 224 pj_oth = NULL; 225 } 226 227 /* first, compute symbolic AP = A_loc*P = A_diag*P_loc + A_off*P_oth */ 228 PetscCall(PetscMalloc1(am + 2, &api)); 229 ptap->api = api; 230 api[0] = 0; 231 232 /* create and initialize a linked list */ 233 PetscCall(PetscLLCondensedCreate(pN, pN, &lnk, &lnkbt)); 234 235 /* Initial FreeSpace size is fill*(nnz(A)+nnz(P)) */ 236 PetscCall(PetscFreeSpaceGet(PetscRealIntMultTruncate(fill, PetscIntSumTruncate(adi[am], PetscIntSumTruncate(aoi[am], pi_loc[pm]))), &free_space)); 237 current_space = free_space; 238 239 MatPreallocateBegin(comm, am, pn, dnz, onz); 240 for (i = 0; i < am; i++) { 241 /* diagonal portion of A */ 242 nzi = adi[i + 1] - adi[i]; 243 for (j = 0; j < nzi; j++) { 244 row = *adj++; 245 pnz = pi_loc[row + 1] - pi_loc[row]; 246 Jptr = pj_loc + pi_loc[row]; 247 /* add non-zero cols of P into the sorted linked list lnk */ 248 PetscCall(PetscLLCondensedAddSorted(pnz, Jptr, lnk, lnkbt)); 249 } 250 /* off-diagonal portion of A */ 251 nzi = aoi[i + 1] - aoi[i]; 252 for (j = 0; j < nzi; j++) { 253 row = *aoj++; 254 pnz = pi_oth[row + 1] - pi_oth[row]; 255 Jptr = pj_oth + pi_oth[row]; 256 PetscCall(PetscLLCondensedAddSorted(pnz, Jptr, lnk, lnkbt)); 257 } 258 /* add possible missing diagonal entry */ 259 if (C->force_diagonals) { 260 j = i + rstart; /* column index */ 261 PetscCall(PetscLLCondensedAddSorted(1, &j, lnk, lnkbt)); 262 } 263 264 apnz = lnk[0]; 265 api[i + 1] = api[i] + apnz; 266 267 /* if free space is not available, double the total space in the list */ 268 if (current_space->local_remaining < apnz) { 269 PetscCall(PetscFreeSpaceGet(PetscIntSumTruncate(apnz, current_space->total_array_size), ¤t_space)); 270 nspacedouble++; 271 } 272 273 /* Copy data into free space, then initialize lnk */ 274 PetscCall(PetscLLCondensedClean(pN, apnz, current_space->array, lnk, lnkbt)); 275 PetscCall(MatPreallocateSet(i + rstart, apnz, current_space->array, dnz, onz)); 276 277 current_space->array += apnz; 278 current_space->local_used += apnz; 279 current_space->local_remaining -= apnz; 280 } 281 282 /* Allocate space for apj, initialize apj, and */ 283 /* destroy list of free space and other temporary array(s) */ 284 PetscCall(PetscMalloc1(api[am] + 1, &ptap->apj)); 285 apj = ptap->apj; 286 PetscCall(PetscFreeSpaceContiguous(&free_space, ptap->apj)); 287 PetscCall(PetscLLDestroy(lnk, lnkbt)); 288 289 /* malloc apa to store dense row A[i,:]*P */ 290 PetscCall(PetscCalloc1(pN, &ptap->apa)); 291 292 /* set and assemble symbolic parallel matrix C */ 293 PetscCall(MatSetSizes(C, am, pn, PETSC_DETERMINE, PETSC_DETERMINE)); 294 PetscCall(MatSetBlockSizesFromMats(C, A, P)); 295 296 PetscCall(MatGetType(A, &mtype)); 297 PetscCall(MatSetType(C, mtype)); 298 PetscCall(MatMPIAIJSetPreallocation(C, 0, dnz, 0, onz)); 299 MatPreallocateEnd(dnz, onz); 300 301 PetscCall(MatSetValues_MPIAIJ_CopyFromCSRFormat_Symbolic(C, apj, api)); 302 PetscCall(MatAssemblyBegin(C, MAT_FINAL_ASSEMBLY)); 303 PetscCall(MatAssemblyEnd(C, MAT_FINAL_ASSEMBLY)); 304 PetscCall(MatSetOption(C, MAT_NEW_NONZERO_LOCATION_ERR, PETSC_TRUE)); 305 306 C->ops->matmultnumeric = MatMatMultNumeric_MPIAIJ_MPIAIJ_nonscalable; 307 C->ops->productnumeric = MatProductNumeric_AB; 308 309 /* attach the supporting struct to C for reuse */ 310 C->product->data = ptap; 311 C->product->destroy = MatDestroy_MPIAIJ_MatMatMult; 312 313 /* set MatInfo */ 314 afill = (PetscReal)api[am] / (adi[am] + aoi[am] + pi_loc[pm] + 1) + 1.e-5; 315 if (afill < 1.0) afill = 1.0; 316 C->info.mallocs = nspacedouble; 317 C->info.fill_ratio_given = fill; 318 C->info.fill_ratio_needed = afill; 319 320 #if defined(PETSC_USE_INFO) 321 if (api[am]) { 322 PetscCall(PetscInfo(C, "Reallocs %" PetscInt_FMT "; Fill ratio: given %g needed %g.\n", nspacedouble, (double)fill, (double)afill)); 323 PetscCall(PetscInfo(C, "Use MatMatMult(A,B,MatReuse,%g,&C) for best performance.;\n", (double)afill)); 324 } else { 325 PetscCall(PetscInfo(C, "Empty matrix product\n")); 326 } 327 #endif 328 PetscFunctionReturn(PETSC_SUCCESS); 329 } 330 331 static PetscErrorCode MatMatMultSymbolic_MPIAIJ_MPIDense(Mat, Mat, PetscReal, Mat); 332 static PetscErrorCode MatMatMultNumeric_MPIAIJ_MPIDense(Mat, Mat, Mat); 333 334 static PetscErrorCode MatProductSetFromOptions_MPIAIJ_MPIDense_AB(Mat C) 335 { 336 Mat_Product *product = C->product; 337 Mat A = product->A, B = product->B; 338 339 PetscFunctionBegin; 340 if (A->cmap->rstart != B->rmap->rstart || A->cmap->rend != B->rmap->rend) 341 SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Matrix local dimensions are incompatible, (%" PetscInt_FMT ", %" PetscInt_FMT ") != (%" PetscInt_FMT ",%" PetscInt_FMT ")", A->cmap->rstart, A->cmap->rend, B->rmap->rstart, B->rmap->rend); 342 343 C->ops->matmultsymbolic = MatMatMultSymbolic_MPIAIJ_MPIDense; 344 C->ops->productsymbolic = MatProductSymbolic_AB; 345 PetscFunctionReturn(PETSC_SUCCESS); 346 } 347 348 static PetscErrorCode MatProductSetFromOptions_MPIAIJ_MPIDense_AtB(Mat C) 349 { 350 Mat_Product *product = C->product; 351 Mat A = product->A, B = product->B; 352 353 PetscFunctionBegin; 354 if (A->rmap->rstart != B->rmap->rstart || A->rmap->rend != B->rmap->rend) 355 SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Matrix local dimensions are incompatible, (%" PetscInt_FMT ", %" PetscInt_FMT ") != (%" PetscInt_FMT ",%" PetscInt_FMT ")", A->rmap->rstart, A->rmap->rend, B->rmap->rstart, B->rmap->rend); 356 357 C->ops->transposematmultsymbolic = MatTransposeMatMultSymbolic_MPIAIJ_MPIDense; 358 C->ops->productsymbolic = MatProductSymbolic_AtB; 359 PetscFunctionReturn(PETSC_SUCCESS); 360 } 361 362 PETSC_INTERN PetscErrorCode MatProductSetFromOptions_MPIAIJ_MPIDense(Mat C) 363 { 364 Mat_Product *product = C->product; 365 366 PetscFunctionBegin; 367 switch (product->type) { 368 case MATPRODUCT_AB: 369 PetscCall(MatProductSetFromOptions_MPIAIJ_MPIDense_AB(C)); 370 break; 371 case MATPRODUCT_AtB: 372 PetscCall(MatProductSetFromOptions_MPIAIJ_MPIDense_AtB(C)); 373 break; 374 default: 375 break; 376 } 377 PetscFunctionReturn(PETSC_SUCCESS); 378 } 379 380 typedef struct { 381 Mat workB, workB1; 382 MPI_Request *rwaits, *swaits; 383 PetscInt nsends, nrecvs; 384 MPI_Datatype *stype, *rtype; 385 PetscInt blda; 386 } MPIAIJ_MPIDense; 387 388 static PetscErrorCode MatMPIAIJ_MPIDenseDestroy(void *ctx) 389 { 390 MPIAIJ_MPIDense *contents = (MPIAIJ_MPIDense *)ctx; 391 PetscInt i; 392 393 PetscFunctionBegin; 394 PetscCall(MatDestroy(&contents->workB)); 395 PetscCall(MatDestroy(&contents->workB1)); 396 for (i = 0; i < contents->nsends; i++) PetscCallMPI(MPI_Type_free(&contents->stype[i])); 397 for (i = 0; i < contents->nrecvs; i++) PetscCallMPI(MPI_Type_free(&contents->rtype[i])); 398 PetscCall(PetscFree4(contents->stype, contents->rtype, contents->rwaits, contents->swaits)); 399 PetscCall(PetscFree(contents)); 400 PetscFunctionReturn(PETSC_SUCCESS); 401 } 402 403 static PetscErrorCode MatMatMultSymbolic_MPIAIJ_MPIDense(Mat A, Mat B, PetscReal fill, Mat C) 404 { 405 Mat_MPIAIJ *aij = (Mat_MPIAIJ *)A->data; 406 PetscInt nz = aij->B->cmap->n, nsends, nrecvs, i, nrows_to, j, blda, m, M, n, N; 407 MPIAIJ_MPIDense *contents; 408 VecScatter ctx = aij->Mvctx; 409 PetscInt Am = A->rmap->n, Bm = B->rmap->n, BN = B->cmap->N, Bbn, Bbn1, bs, nrows_from, numBb; 410 MPI_Comm comm; 411 MPI_Datatype type1, *stype, *rtype; 412 const PetscInt *sindices, *sstarts, *rstarts; 413 PetscMPIInt *disp; 414 PetscBool cisdense; 415 416 PetscFunctionBegin; 417 MatCheckProduct(C, 4); 418 PetscCheck(!C->product->data, PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Product data not empty"); 419 PetscCall(PetscObjectGetComm((PetscObject)A, &comm)); 420 PetscCall(PetscObjectBaseTypeCompare((PetscObject)C, MATMPIDENSE, &cisdense)); 421 if (!cisdense) PetscCall(MatSetType(C, ((PetscObject)B)->type_name)); 422 PetscCall(MatGetLocalSize(C, &m, &n)); 423 PetscCall(MatGetSize(C, &M, &N)); 424 if (m == PETSC_DECIDE || n == PETSC_DECIDE || M == PETSC_DECIDE || N == PETSC_DECIDE) PetscCall(MatSetSizes(C, Am, B->cmap->n, A->rmap->N, BN)); 425 PetscCall(MatSetBlockSizesFromMats(C, A, B)); 426 PetscCall(MatSetUp(C)); 427 PetscCall(MatDenseGetLDA(B, &blda)); 428 PetscCall(PetscNew(&contents)); 429 430 PetscCall(VecScatterGetRemote_Private(ctx, PETSC_TRUE /*send*/, &nsends, &sstarts, &sindices, NULL, NULL)); 431 PetscCall(VecScatterGetRemoteOrdered_Private(ctx, PETSC_FALSE /*recv*/, &nrecvs, &rstarts, NULL, NULL, NULL)); 432 433 /* Create column block of B and C for memory scalability when BN is too large */ 434 /* Estimate Bbn, column size of Bb */ 435 if (nz) { 436 Bbn1 = 2 * Am * BN / nz; 437 if (!Bbn1) Bbn1 = 1; 438 } else Bbn1 = BN; 439 440 bs = PetscAbs(B->cmap->bs); 441 Bbn1 = Bbn1 / bs * bs; /* Bbn1 is a multiple of bs */ 442 if (Bbn1 > BN) Bbn1 = BN; 443 PetscCall(MPIU_Allreduce(&Bbn1, &Bbn, 1, MPIU_INT, MPI_MAX, comm)); 444 445 /* Enable runtime option for Bbn */ 446 PetscOptionsBegin(comm, ((PetscObject)C)->prefix, "MatMatMult", "Mat"); 447 PetscCall(PetscOptionsInt("-matmatmult_Bbn", "Number of columns in Bb", "MatMatMult", Bbn, &Bbn, NULL)); 448 PetscOptionsEnd(); 449 Bbn = PetscMin(Bbn, BN); 450 451 if (Bbn > 0 && Bbn < BN) { 452 numBb = BN / Bbn; 453 Bbn1 = BN - numBb * Bbn; 454 } else numBb = 0; 455 456 if (numBb) { 457 PetscCall(PetscInfo(C, "use Bb, BN=%" PetscInt_FMT ", Bbn=%" PetscInt_FMT "; numBb=%" PetscInt_FMT "\n", BN, Bbn, numBb)); 458 if (Bbn1) { /* Create workB1 for the remaining columns */ 459 PetscCall(PetscInfo(C, "use Bb1, BN=%" PetscInt_FMT ", Bbn1=%" PetscInt_FMT "\n", BN, Bbn1)); 460 /* Create work matrix used to store off processor rows of B needed for local product */ 461 PetscCall(MatCreateSeqDense(PETSC_COMM_SELF, nz, Bbn1, NULL, &contents->workB1)); 462 } else contents->workB1 = NULL; 463 } 464 465 /* Create work matrix used to store off processor rows of B needed for local product */ 466 PetscCall(MatCreateSeqDense(PETSC_COMM_SELF, nz, Bbn, NULL, &contents->workB)); 467 468 /* Use MPI derived data type to reduce memory required by the send/recv buffers */ 469 PetscCall(PetscMalloc4(nsends, &stype, nrecvs, &rtype, nrecvs, &contents->rwaits, nsends, &contents->swaits)); 470 contents->stype = stype; 471 contents->nsends = nsends; 472 473 contents->rtype = rtype; 474 contents->nrecvs = nrecvs; 475 contents->blda = blda; 476 477 PetscCall(PetscMalloc1(Bm + 1, &disp)); 478 for (i = 0; i < nsends; i++) { 479 nrows_to = sstarts[i + 1] - sstarts[i]; 480 for (j = 0; j < nrows_to; j++) disp[j] = sindices[sstarts[i] + j]; /* rowB to be sent */ 481 PetscCallMPI(MPI_Type_create_indexed_block(nrows_to, 1, disp, MPIU_SCALAR, &type1)); 482 PetscCallMPI(MPI_Type_create_resized(type1, 0, blda * sizeof(PetscScalar), &stype[i])); 483 PetscCallMPI(MPI_Type_commit(&stype[i])); 484 PetscCallMPI(MPI_Type_free(&type1)); 485 } 486 487 for (i = 0; i < nrecvs; i++) { 488 /* received values from a process form a (nrows_from x Bbn) row block in workB (column-wise) */ 489 nrows_from = rstarts[i + 1] - rstarts[i]; 490 disp[0] = 0; 491 PetscCallMPI(MPI_Type_create_indexed_block(1, nrows_from, disp, MPIU_SCALAR, &type1)); 492 PetscCallMPI(MPI_Type_create_resized(type1, 0, nz * sizeof(PetscScalar), &rtype[i])); 493 PetscCallMPI(MPI_Type_commit(&rtype[i])); 494 PetscCallMPI(MPI_Type_free(&type1)); 495 } 496 497 PetscCall(PetscFree(disp)); 498 PetscCall(VecScatterRestoreRemote_Private(ctx, PETSC_TRUE /*send*/, &nsends, &sstarts, &sindices, NULL, NULL)); 499 PetscCall(VecScatterRestoreRemoteOrdered_Private(ctx, PETSC_FALSE /*recv*/, &nrecvs, &rstarts, NULL, NULL, NULL)); 500 PetscCall(MatSetOption(C, MAT_NO_OFF_PROC_ENTRIES, PETSC_TRUE)); 501 PetscCall(MatAssemblyBegin(C, MAT_FINAL_ASSEMBLY)); 502 PetscCall(MatAssemblyEnd(C, MAT_FINAL_ASSEMBLY)); 503 PetscCall(MatSetOption(C, MAT_NEW_NONZERO_LOCATION_ERR, PETSC_TRUE)); 504 505 C->product->data = contents; 506 C->product->destroy = MatMPIAIJ_MPIDenseDestroy; 507 C->ops->matmultnumeric = MatMatMultNumeric_MPIAIJ_MPIDense; 508 PetscFunctionReturn(PETSC_SUCCESS); 509 } 510 511 PETSC_INTERN PetscErrorCode MatMatMultNumericAdd_SeqAIJ_SeqDense(Mat, Mat, Mat, const PetscBool); 512 513 /* 514 Performs an efficient scatter on the rows of B needed by this process; this is 515 a modification of the VecScatterBegin_() routines. 516 517 Input: If Bbidx = 0, uses B = Bb, else B = Bb1, see MatMatMultSymbolic_MPIAIJ_MPIDense() 518 */ 519 520 static PetscErrorCode MatMPIDenseScatter(Mat A, Mat B, PetscInt Bbidx, Mat C, Mat *outworkB) 521 { 522 Mat_MPIAIJ *aij = (Mat_MPIAIJ *)A->data; 523 const PetscScalar *b; 524 PetscScalar *rvalues; 525 VecScatter ctx = aij->Mvctx; 526 const PetscInt *sindices, *sstarts, *rstarts; 527 const PetscMPIInt *sprocs, *rprocs; 528 PetscInt i, nsends, nrecvs; 529 MPI_Request *swaits, *rwaits; 530 MPI_Comm comm; 531 PetscMPIInt tag = ((PetscObject)ctx)->tag, ncols = B->cmap->N, nrows = aij->B->cmap->n, nsends_mpi, nrecvs_mpi; 532 MPIAIJ_MPIDense *contents; 533 Mat workB; 534 MPI_Datatype *stype, *rtype; 535 PetscInt blda; 536 537 PetscFunctionBegin; 538 MatCheckProduct(C, 4); 539 PetscCheck(C->product->data, PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Product data empty"); 540 contents = (MPIAIJ_MPIDense *)C->product->data; 541 PetscCall(VecScatterGetRemote_Private(ctx, PETSC_TRUE /*send*/, &nsends, &sstarts, &sindices, &sprocs, NULL /*bs*/)); 542 PetscCall(VecScatterGetRemoteOrdered_Private(ctx, PETSC_FALSE /*recv*/, &nrecvs, &rstarts, NULL, &rprocs, NULL /*bs*/)); 543 PetscCall(PetscMPIIntCast(nsends, &nsends_mpi)); 544 PetscCall(PetscMPIIntCast(nrecvs, &nrecvs_mpi)); 545 if (Bbidx == 0) workB = *outworkB = contents->workB; 546 else workB = *outworkB = contents->workB1; 547 PetscCheck(nrows == workB->rmap->n, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Number of rows of workB %" PetscInt_FMT " not equal to columns of aij->B %d", workB->cmap->n, nrows); 548 swaits = contents->swaits; 549 rwaits = contents->rwaits; 550 551 PetscCall(MatDenseGetArrayRead(B, &b)); 552 PetscCall(MatDenseGetLDA(B, &blda)); 553 PetscCheck(blda == contents->blda, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Cannot reuse an input matrix with lda %" PetscInt_FMT " != %" PetscInt_FMT, blda, contents->blda); 554 PetscCall(MatDenseGetArray(workB, &rvalues)); 555 556 /* Post recv, use MPI derived data type to save memory */ 557 PetscCall(PetscObjectGetComm((PetscObject)C, &comm)); 558 rtype = contents->rtype; 559 for (i = 0; i < nrecvs; i++) PetscCallMPI(MPI_Irecv(rvalues + (rstarts[i] - rstarts[0]), ncols, rtype[i], rprocs[i], tag, comm, rwaits + i)); 560 561 stype = contents->stype; 562 for (i = 0; i < nsends; i++) PetscCallMPI(MPI_Isend(b, ncols, stype[i], sprocs[i], tag, comm, swaits + i)); 563 564 if (nrecvs) PetscCallMPI(MPI_Waitall(nrecvs_mpi, rwaits, MPI_STATUSES_IGNORE)); 565 if (nsends) PetscCallMPI(MPI_Waitall(nsends_mpi, swaits, MPI_STATUSES_IGNORE)); 566 567 PetscCall(VecScatterRestoreRemote_Private(ctx, PETSC_TRUE /*send*/, &nsends, &sstarts, &sindices, &sprocs, NULL)); 568 PetscCall(VecScatterRestoreRemoteOrdered_Private(ctx, PETSC_FALSE /*recv*/, &nrecvs, &rstarts, NULL, &rprocs, NULL)); 569 PetscCall(MatDenseRestoreArrayRead(B, &b)); 570 PetscCall(MatDenseRestoreArray(workB, &rvalues)); 571 PetscFunctionReturn(PETSC_SUCCESS); 572 } 573 574 static PetscErrorCode MatMatMultNumeric_MPIAIJ_MPIDense(Mat A, Mat B, Mat C) 575 { 576 Mat_MPIAIJ *aij = (Mat_MPIAIJ *)A->data; 577 Mat_MPIDense *bdense = (Mat_MPIDense *)B->data; 578 Mat_MPIDense *cdense = (Mat_MPIDense *)C->data; 579 Mat workB; 580 MPIAIJ_MPIDense *contents; 581 582 PetscFunctionBegin; 583 MatCheckProduct(C, 3); 584 PetscCheck(C->product->data, PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Product data empty"); 585 contents = (MPIAIJ_MPIDense *)C->product->data; 586 /* diagonal block of A times all local rows of B */ 587 /* TODO: this calls a symbolic multiplication every time, which could be avoided */ 588 PetscCall(MatMatMult(aij->A, bdense->A, MAT_REUSE_MATRIX, PETSC_DEFAULT, &cdense->A)); 589 if (contents->workB->cmap->n == B->cmap->N) { 590 /* get off processor parts of B needed to complete C=A*B */ 591 PetscCall(MatMPIDenseScatter(A, B, 0, C, &workB)); 592 593 /* off-diagonal block of A times nonlocal rows of B */ 594 PetscCall(MatMatMultNumericAdd_SeqAIJ_SeqDense(aij->B, workB, cdense->A, PETSC_TRUE)); 595 } else { 596 Mat Bb, Cb; 597 PetscInt BN = B->cmap->N, n = contents->workB->cmap->n, i; 598 PetscBool ccpu; 599 600 PetscCheck(n > 0, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Column block size %" PetscInt_FMT " must be positive", n); 601 /* Prevent from unneeded copies back and forth from the GPU 602 when getting and restoring the submatrix 603 We need a proper GPU code for AIJ * dense in parallel */ 604 PetscCall(MatBoundToCPU(C, &ccpu)); 605 PetscCall(MatBindToCPU(C, PETSC_TRUE)); 606 for (i = 0; i < BN; i += n) { 607 PetscCall(MatDenseGetSubMatrix(B, PETSC_DECIDE, PETSC_DECIDE, i, PetscMin(i + n, BN), &Bb)); 608 PetscCall(MatDenseGetSubMatrix(C, PETSC_DECIDE, PETSC_DECIDE, i, PetscMin(i + n, BN), &Cb)); 609 610 /* get off processor parts of B needed to complete C=A*B */ 611 PetscCall(MatMPIDenseScatter(A, Bb, (i + n) > BN, C, &workB)); 612 613 /* off-diagonal block of A times nonlocal rows of B */ 614 cdense = (Mat_MPIDense *)Cb->data; 615 PetscCall(MatMatMultNumericAdd_SeqAIJ_SeqDense(aij->B, workB, cdense->A, PETSC_TRUE)); 616 PetscCall(MatDenseRestoreSubMatrix(B, &Bb)); 617 PetscCall(MatDenseRestoreSubMatrix(C, &Cb)); 618 } 619 PetscCall(MatBindToCPU(C, ccpu)); 620 } 621 PetscFunctionReturn(PETSC_SUCCESS); 622 } 623 624 PetscErrorCode MatMatMultNumeric_MPIAIJ_MPIAIJ(Mat A, Mat P, Mat C) 625 { 626 Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data, *c = (Mat_MPIAIJ *)C->data; 627 Mat_SeqAIJ *ad = (Mat_SeqAIJ *)(a->A)->data, *ao = (Mat_SeqAIJ *)(a->B)->data; 628 Mat_SeqAIJ *cd = (Mat_SeqAIJ *)(c->A)->data, *co = (Mat_SeqAIJ *)(c->B)->data; 629 PetscInt *adi = ad->i, *adj, *aoi = ao->i, *aoj; 630 PetscScalar *ada, *aoa, *cda = cd->a, *coa = co->a; 631 Mat_SeqAIJ *p_loc, *p_oth; 632 PetscInt *pi_loc, *pj_loc, *pi_oth, *pj_oth, *pj; 633 PetscScalar *pa_loc, *pa_oth, *pa, valtmp, *ca; 634 PetscInt cm = C->rmap->n, anz, pnz; 635 Mat_APMPI *ptap; 636 PetscScalar *apa_sparse; 637 const PetscScalar *dummy; 638 PetscInt *api, *apj, *apJ, i, j, k, row; 639 PetscInt cstart = C->cmap->rstart; 640 PetscInt cdnz, conz, k0, k1, nextp; 641 MPI_Comm comm; 642 PetscMPIInt size; 643 644 PetscFunctionBegin; 645 MatCheckProduct(C, 3); 646 ptap = (Mat_APMPI *)C->product->data; 647 PetscCheck(ptap, PetscObjectComm((PetscObject)C), PETSC_ERR_ARG_WRONGSTATE, "PtAP cannot be computed. Missing data"); 648 PetscCall(PetscObjectGetComm((PetscObject)C, &comm)); 649 PetscCallMPI(MPI_Comm_size(comm, &size)); 650 PetscCheck(ptap->P_oth || size <= 1, PetscObjectComm((PetscObject)C), PETSC_ERR_ARG_WRONGSTATE, "AP cannot be reused. Do not call MatProductClear()"); 651 652 /* flag CPU mask for C */ 653 #if defined(PETSC_HAVE_DEVICE) 654 if (C->offloadmask != PETSC_OFFLOAD_UNALLOCATED) C->offloadmask = PETSC_OFFLOAD_CPU; 655 if (c->A->offloadmask != PETSC_OFFLOAD_UNALLOCATED) c->A->offloadmask = PETSC_OFFLOAD_CPU; 656 if (c->B->offloadmask != PETSC_OFFLOAD_UNALLOCATED) c->B->offloadmask = PETSC_OFFLOAD_CPU; 657 #endif 658 apa_sparse = ptap->apa; 659 660 /* 1) get P_oth = ptap->P_oth and P_loc = ptap->P_loc */ 661 /* update numerical values of P_oth and P_loc */ 662 PetscCall(MatGetBrowsOfAoCols_MPIAIJ(A, P, MAT_REUSE_MATRIX, &ptap->startsj_s, &ptap->startsj_r, &ptap->bufa, &ptap->P_oth)); 663 PetscCall(MatMPIAIJGetLocalMat(P, MAT_REUSE_MATRIX, &ptap->P_loc)); 664 665 /* 2) compute numeric C_loc = A_loc*P = Ad*P_loc + Ao*P_oth */ 666 /* get data from symbolic products */ 667 p_loc = (Mat_SeqAIJ *)(ptap->P_loc)->data; 668 pi_loc = p_loc->i; 669 pj_loc = p_loc->j; 670 pa_loc = p_loc->a; 671 if (size > 1) { 672 p_oth = (Mat_SeqAIJ *)(ptap->P_oth)->data; 673 pi_oth = p_oth->i; 674 pj_oth = p_oth->j; 675 pa_oth = p_oth->a; 676 } else { 677 p_oth = NULL; 678 pi_oth = NULL; 679 pj_oth = NULL; 680 pa_oth = NULL; 681 } 682 683 /* trigger copy to CPU */ 684 PetscCall(MatSeqAIJGetArrayRead(a->A, &dummy)); 685 PetscCall(MatSeqAIJRestoreArrayRead(a->A, &dummy)); 686 PetscCall(MatSeqAIJGetArrayRead(a->B, &dummy)); 687 PetscCall(MatSeqAIJRestoreArrayRead(a->B, &dummy)); 688 api = ptap->api; 689 apj = ptap->apj; 690 for (i = 0; i < cm; i++) { 691 apJ = apj + api[i]; 692 693 /* diagonal portion of A */ 694 anz = adi[i + 1] - adi[i]; 695 adj = ad->j + adi[i]; 696 ada = ad->a + adi[i]; 697 for (j = 0; j < anz; j++) { 698 row = adj[j]; 699 pnz = pi_loc[row + 1] - pi_loc[row]; 700 pj = pj_loc + pi_loc[row]; 701 pa = pa_loc + pi_loc[row]; 702 /* perform sparse axpy */ 703 valtmp = ada[j]; 704 nextp = 0; 705 for (k = 0; nextp < pnz; k++) { 706 if (apJ[k] == pj[nextp]) { /* column of AP == column of P */ 707 apa_sparse[k] += valtmp * pa[nextp++]; 708 } 709 } 710 PetscCall(PetscLogFlops(2.0 * pnz)); 711 } 712 713 /* off-diagonal portion of A */ 714 anz = aoi[i + 1] - aoi[i]; 715 aoj = ao->j + aoi[i]; 716 aoa = ao->a + aoi[i]; 717 for (j = 0; j < anz; j++) { 718 row = aoj[j]; 719 pnz = pi_oth[row + 1] - pi_oth[row]; 720 pj = pj_oth + pi_oth[row]; 721 pa = pa_oth + pi_oth[row]; 722 /* perform sparse axpy */ 723 valtmp = aoa[j]; 724 nextp = 0; 725 for (k = 0; nextp < pnz; k++) { 726 if (apJ[k] == pj[nextp]) { /* column of AP == column of P */ 727 apa_sparse[k] += valtmp * pa[nextp++]; 728 } 729 } 730 PetscCall(PetscLogFlops(2.0 * pnz)); 731 } 732 733 /* set values in C */ 734 cdnz = cd->i[i + 1] - cd->i[i]; 735 conz = co->i[i + 1] - co->i[i]; 736 737 /* 1st off-diagonal part of C */ 738 ca = coa + co->i[i]; 739 k = 0; 740 for (k0 = 0; k0 < conz; k0++) { 741 if (apJ[k] >= cstart) break; 742 ca[k0] = apa_sparse[k]; 743 apa_sparse[k] = 0.0; 744 k++; 745 } 746 747 /* diagonal part of C */ 748 ca = cda + cd->i[i]; 749 for (k1 = 0; k1 < cdnz; k1++) { 750 ca[k1] = apa_sparse[k]; 751 apa_sparse[k] = 0.0; 752 k++; 753 } 754 755 /* 2nd off-diagonal part of C */ 756 ca = coa + co->i[i]; 757 for (; k0 < conz; k0++) { 758 ca[k0] = apa_sparse[k]; 759 apa_sparse[k] = 0.0; 760 k++; 761 } 762 } 763 PetscCall(MatAssemblyBegin(C, MAT_FINAL_ASSEMBLY)); 764 PetscCall(MatAssemblyEnd(C, MAT_FINAL_ASSEMBLY)); 765 PetscFunctionReturn(PETSC_SUCCESS); 766 } 767 768 /* same as MatMatMultSymbolic_MPIAIJ_MPIAIJ_nonscalable(), except using LLCondensed to avoid O(BN) memory requirement */ 769 PetscErrorCode MatMatMultSymbolic_MPIAIJ_MPIAIJ(Mat A, Mat P, PetscReal fill, Mat C) 770 { 771 MPI_Comm comm; 772 PetscMPIInt size; 773 Mat_APMPI *ptap; 774 PetscFreeSpaceList free_space = NULL, current_space = NULL; 775 Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data; 776 Mat_SeqAIJ *ad = (Mat_SeqAIJ *)(a->A)->data, *ao = (Mat_SeqAIJ *)(a->B)->data, *p_loc, *p_oth; 777 PetscInt *pi_loc, *pj_loc, *pi_oth, *pj_oth, *dnz, *onz; 778 PetscInt *adi = ad->i, *adj = ad->j, *aoi = ao->i, *aoj = ao->j, rstart = A->rmap->rstart; 779 PetscInt i, pnz, row, *api, *apj, *Jptr, apnz, nspacedouble = 0, j, nzi, *lnk, apnz_max = 1; 780 PetscInt am = A->rmap->n, pn = P->cmap->n, pm = P->rmap->n, lsize = pn + 20; 781 PetscReal afill; 782 MatType mtype; 783 784 PetscFunctionBegin; 785 MatCheckProduct(C, 4); 786 PetscCheck(!C->product->data, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Extra product struct not empty"); 787 PetscCall(PetscObjectGetComm((PetscObject)A, &comm)); 788 PetscCallMPI(MPI_Comm_size(comm, &size)); 789 790 /* create struct Mat_APMPI and attached it to C later */ 791 PetscCall(PetscNew(&ptap)); 792 793 /* get P_oth by taking rows of P (= non-zero cols of local A) from other processors */ 794 PetscCall(MatGetBrowsOfAoCols_MPIAIJ(A, P, MAT_INITIAL_MATRIX, &ptap->startsj_s, &ptap->startsj_r, &ptap->bufa, &ptap->P_oth)); 795 796 /* get P_loc by taking all local rows of P */ 797 PetscCall(MatMPIAIJGetLocalMat(P, MAT_INITIAL_MATRIX, &ptap->P_loc)); 798 799 p_loc = (Mat_SeqAIJ *)(ptap->P_loc)->data; 800 pi_loc = p_loc->i; 801 pj_loc = p_loc->j; 802 if (size > 1) { 803 p_oth = (Mat_SeqAIJ *)(ptap->P_oth)->data; 804 pi_oth = p_oth->i; 805 pj_oth = p_oth->j; 806 } else { 807 p_oth = NULL; 808 pi_oth = NULL; 809 pj_oth = NULL; 810 } 811 812 /* first, compute symbolic AP = A_loc*P = A_diag*P_loc + A_off*P_oth */ 813 PetscCall(PetscMalloc1(am + 2, &api)); 814 ptap->api = api; 815 api[0] = 0; 816 817 PetscCall(PetscLLCondensedCreate_Scalable(lsize, &lnk)); 818 819 /* Initial FreeSpace size is fill*(nnz(A)+nnz(P)) */ 820 PetscCall(PetscFreeSpaceGet(PetscRealIntMultTruncate(fill, PetscIntSumTruncate(adi[am], PetscIntSumTruncate(aoi[am], pi_loc[pm]))), &free_space)); 821 current_space = free_space; 822 MatPreallocateBegin(comm, am, pn, dnz, onz); 823 for (i = 0; i < am; i++) { 824 /* diagonal portion of A */ 825 nzi = adi[i + 1] - adi[i]; 826 for (j = 0; j < nzi; j++) { 827 row = *adj++; 828 pnz = pi_loc[row + 1] - pi_loc[row]; 829 Jptr = pj_loc + pi_loc[row]; 830 /* Expand list if it is not long enough */ 831 if (pnz + apnz_max > lsize) { 832 lsize = pnz + apnz_max; 833 PetscCall(PetscLLCondensedExpand_Scalable(lsize, &lnk)); 834 } 835 /* add non-zero cols of P into the sorted linked list lnk */ 836 PetscCall(PetscLLCondensedAddSorted_Scalable(pnz, Jptr, lnk)); 837 apnz = *lnk; /* The first element in the list is the number of items in the list */ 838 api[i + 1] = api[i] + apnz; 839 if (apnz > apnz_max) apnz_max = apnz + 1; /* '1' for diagonal entry */ 840 } 841 /* off-diagonal portion of A */ 842 nzi = aoi[i + 1] - aoi[i]; 843 for (j = 0; j < nzi; j++) { 844 row = *aoj++; 845 pnz = pi_oth[row + 1] - pi_oth[row]; 846 Jptr = pj_oth + pi_oth[row]; 847 /* Expand list if it is not long enough */ 848 if (pnz + apnz_max > lsize) { 849 lsize = pnz + apnz_max; 850 PetscCall(PetscLLCondensedExpand_Scalable(lsize, &lnk)); 851 } 852 /* add non-zero cols of P into the sorted linked list lnk */ 853 PetscCall(PetscLLCondensedAddSorted_Scalable(pnz, Jptr, lnk)); 854 apnz = *lnk; /* The first element in the list is the number of items in the list */ 855 api[i + 1] = api[i] + apnz; 856 if (apnz > apnz_max) apnz_max = apnz + 1; /* '1' for diagonal entry */ 857 } 858 859 /* add missing diagonal entry */ 860 if (C->force_diagonals) { 861 j = i + rstart; /* column index */ 862 PetscCall(PetscLLCondensedAddSorted_Scalable(1, &j, lnk)); 863 } 864 865 apnz = *lnk; 866 api[i + 1] = api[i] + apnz; 867 if (apnz > apnz_max) apnz_max = apnz; 868 869 /* if free space is not available, double the total space in the list */ 870 if (current_space->local_remaining < apnz) { 871 PetscCall(PetscFreeSpaceGet(PetscIntSumTruncate(apnz, current_space->total_array_size), ¤t_space)); 872 nspacedouble++; 873 } 874 875 /* Copy data into free space, then initialize lnk */ 876 PetscCall(PetscLLCondensedClean_Scalable(apnz, current_space->array, lnk)); 877 PetscCall(MatPreallocateSet(i + rstart, apnz, current_space->array, dnz, onz)); 878 879 current_space->array += apnz; 880 current_space->local_used += apnz; 881 current_space->local_remaining -= apnz; 882 } 883 884 /* Allocate space for apj, initialize apj, and */ 885 /* destroy list of free space and other temporary array(s) */ 886 PetscCall(PetscMalloc1(api[am] + 1, &ptap->apj)); 887 apj = ptap->apj; 888 PetscCall(PetscFreeSpaceContiguous(&free_space, ptap->apj)); 889 PetscCall(PetscLLCondensedDestroy_Scalable(lnk)); 890 891 /* create and assemble symbolic parallel matrix C */ 892 PetscCall(MatSetSizes(C, am, pn, PETSC_DETERMINE, PETSC_DETERMINE)); 893 PetscCall(MatSetBlockSizesFromMats(C, A, P)); 894 PetscCall(MatGetType(A, &mtype)); 895 PetscCall(MatSetType(C, mtype)); 896 PetscCall(MatMPIAIJSetPreallocation(C, 0, dnz, 0, onz)); 897 MatPreallocateEnd(dnz, onz); 898 899 /* malloc apa for assembly C */ 900 PetscCall(PetscCalloc1(apnz_max, &ptap->apa)); 901 902 PetscCall(MatSetValues_MPIAIJ_CopyFromCSRFormat_Symbolic(C, apj, api)); 903 PetscCall(MatAssemblyBegin(C, MAT_FINAL_ASSEMBLY)); 904 PetscCall(MatAssemblyEnd(C, MAT_FINAL_ASSEMBLY)); 905 PetscCall(MatSetOption(C, MAT_NEW_NONZERO_LOCATION_ERR, PETSC_TRUE)); 906 907 C->ops->matmultnumeric = MatMatMultNumeric_MPIAIJ_MPIAIJ; 908 C->ops->productnumeric = MatProductNumeric_AB; 909 910 /* attach the supporting struct to C for reuse */ 911 C->product->data = ptap; 912 C->product->destroy = MatDestroy_MPIAIJ_MatMatMult; 913 914 /* set MatInfo */ 915 afill = (PetscReal)api[am] / (adi[am] + aoi[am] + pi_loc[pm] + 1) + 1.e-5; 916 if (afill < 1.0) afill = 1.0; 917 C->info.mallocs = nspacedouble; 918 C->info.fill_ratio_given = fill; 919 C->info.fill_ratio_needed = afill; 920 921 #if defined(PETSC_USE_INFO) 922 if (api[am]) { 923 PetscCall(PetscInfo(C, "Reallocs %" PetscInt_FMT "; Fill ratio: given %g needed %g.\n", nspacedouble, (double)fill, (double)afill)); 924 PetscCall(PetscInfo(C, "Use MatMatMult(A,B,MatReuse,%g,&C) for best performance.;\n", (double)afill)); 925 } else { 926 PetscCall(PetscInfo(C, "Empty matrix product\n")); 927 } 928 #endif 929 PetscFunctionReturn(PETSC_SUCCESS); 930 } 931 932 /* This function is needed for the seqMPI matrix-matrix multiplication. */ 933 /* Three input arrays are merged to one output array. The size of the */ 934 /* output array is also output. Duplicate entries only show up once. */ 935 static void Merge3SortedArrays(PetscInt size1, PetscInt *in1, PetscInt size2, PetscInt *in2, PetscInt size3, PetscInt *in3, PetscInt *size4, PetscInt *out) 936 { 937 int i = 0, j = 0, k = 0, l = 0; 938 939 /* Traverse all three arrays */ 940 while (i < size1 && j < size2 && k < size3) { 941 if (in1[i] < in2[j] && in1[i] < in3[k]) { 942 out[l++] = in1[i++]; 943 } else if (in2[j] < in1[i] && in2[j] < in3[k]) { 944 out[l++] = in2[j++]; 945 } else if (in3[k] < in1[i] && in3[k] < in2[j]) { 946 out[l++] = in3[k++]; 947 } else if (in1[i] == in2[j] && in1[i] < in3[k]) { 948 out[l++] = in1[i]; 949 i++, j++; 950 } else if (in1[i] == in3[k] && in1[i] < in2[j]) { 951 out[l++] = in1[i]; 952 i++, k++; 953 } else if (in3[k] == in2[j] && in2[j] < in1[i]) { 954 out[l++] = in2[j]; 955 k++, j++; 956 } else if (in1[i] == in2[j] && in1[i] == in3[k]) { 957 out[l++] = in1[i]; 958 i++, j++, k++; 959 } 960 } 961 962 /* Traverse two remaining arrays */ 963 while (i < size1 && j < size2) { 964 if (in1[i] < in2[j]) { 965 out[l++] = in1[i++]; 966 } else if (in1[i] > in2[j]) { 967 out[l++] = in2[j++]; 968 } else { 969 out[l++] = in1[i]; 970 i++, j++; 971 } 972 } 973 974 while (i < size1 && k < size3) { 975 if (in1[i] < in3[k]) { 976 out[l++] = in1[i++]; 977 } else if (in1[i] > in3[k]) { 978 out[l++] = in3[k++]; 979 } else { 980 out[l++] = in1[i]; 981 i++, k++; 982 } 983 } 984 985 while (k < size3 && j < size2) { 986 if (in3[k] < in2[j]) { 987 out[l++] = in3[k++]; 988 } else if (in3[k] > in2[j]) { 989 out[l++] = in2[j++]; 990 } else { 991 out[l++] = in3[k]; 992 k++, j++; 993 } 994 } 995 996 /* Traverse one remaining array */ 997 while (i < size1) out[l++] = in1[i++]; 998 while (j < size2) out[l++] = in2[j++]; 999 while (k < size3) out[l++] = in3[k++]; 1000 1001 *size4 = l; 1002 } 1003 1004 /* This matrix-matrix multiplication algorithm divides the multiplication into three multiplications and */ 1005 /* adds up the products. Two of these three multiplications are performed with existing (sequential) */ 1006 /* matrix-matrix multiplications. */ 1007 PetscErrorCode MatMatMultSymbolic_MPIAIJ_MPIAIJ_seqMPI(Mat A, Mat P, PetscReal fill, Mat C) 1008 { 1009 MPI_Comm comm; 1010 PetscMPIInt size; 1011 Mat_APMPI *ptap; 1012 PetscFreeSpaceList free_space_diag = NULL, current_space = NULL; 1013 Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data; 1014 Mat_SeqAIJ *ad = (Mat_SeqAIJ *)(a->A)->data, *ao = (Mat_SeqAIJ *)(a->B)->data, *p_loc; 1015 Mat_MPIAIJ *p = (Mat_MPIAIJ *)P->data; 1016 Mat_SeqAIJ *adpd_seq, *p_off, *aopoth_seq; 1017 PetscInt adponz, adpdnz; 1018 PetscInt *pi_loc, *dnz, *onz; 1019 PetscInt *adi = ad->i, *adj = ad->j, *aoi = ao->i, rstart = A->rmap->rstart; 1020 PetscInt *lnk, i, i1 = 0, pnz, row, *adpoi, *adpoj, *api, *adpoJ, *aopJ, *apJ, *Jptr, aopnz, nspacedouble = 0, j, nzi, *apj, apnz, *adpdi, *adpdj, *adpdJ, *poff_i, *poff_j, *j_temp, *aopothi, *aopothj; 1021 PetscInt am = A->rmap->n, pN = P->cmap->N, pn = P->cmap->n, pm = P->rmap->n, p_colstart, p_colend; 1022 PetscBT lnkbt; 1023 PetscReal afill; 1024 PetscMPIInt rank; 1025 Mat adpd, aopoth; 1026 MatType mtype; 1027 const char *prefix; 1028 1029 PetscFunctionBegin; 1030 MatCheckProduct(C, 4); 1031 PetscCheck(!C->product->data, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Extra product struct not empty"); 1032 PetscCall(PetscObjectGetComm((PetscObject)A, &comm)); 1033 PetscCallMPI(MPI_Comm_size(comm, &size)); 1034 PetscCallMPI(MPI_Comm_rank(comm, &rank)); 1035 PetscCall(MatGetOwnershipRangeColumn(P, &p_colstart, &p_colend)); 1036 1037 /* create struct Mat_APMPI and attached it to C later */ 1038 PetscCall(PetscNew(&ptap)); 1039 1040 /* get P_oth by taking rows of P (= non-zero cols of local A) from other processors */ 1041 PetscCall(MatGetBrowsOfAoCols_MPIAIJ(A, P, MAT_INITIAL_MATRIX, &ptap->startsj_s, &ptap->startsj_r, &ptap->bufa, &ptap->P_oth)); 1042 1043 /* get P_loc by taking all local rows of P */ 1044 PetscCall(MatMPIAIJGetLocalMat(P, MAT_INITIAL_MATRIX, &ptap->P_loc)); 1045 1046 p_loc = (Mat_SeqAIJ *)(ptap->P_loc)->data; 1047 pi_loc = p_loc->i; 1048 1049 /* Allocate memory for the i arrays of the matrices A*P, A_diag*P_off and A_offd * P */ 1050 PetscCall(PetscMalloc1(am + 2, &api)); 1051 PetscCall(PetscMalloc1(am + 2, &adpoi)); 1052 1053 adpoi[0] = 0; 1054 ptap->api = api; 1055 api[0] = 0; 1056 1057 /* create and initialize a linked list, will be used for both A_diag * P_loc_off and A_offd * P_oth */ 1058 PetscCall(PetscLLCondensedCreate(pN, pN, &lnk, &lnkbt)); 1059 MatPreallocateBegin(comm, am, pn, dnz, onz); 1060 1061 /* Symbolic calc of A_loc_diag * P_loc_diag */ 1062 PetscCall(MatGetOptionsPrefix(A, &prefix)); 1063 PetscCall(MatProductCreate(a->A, p->A, NULL, &adpd)); 1064 PetscCall(MatGetOptionsPrefix(A, &prefix)); 1065 PetscCall(MatSetOptionsPrefix(adpd, prefix)); 1066 PetscCall(MatAppendOptionsPrefix(adpd, "inner_diag_")); 1067 1068 PetscCall(MatProductSetType(adpd, MATPRODUCT_AB)); 1069 PetscCall(MatProductSetAlgorithm(adpd, "sorted")); 1070 PetscCall(MatProductSetFill(adpd, fill)); 1071 PetscCall(MatProductSetFromOptions(adpd)); 1072 1073 adpd->force_diagonals = C->force_diagonals; 1074 PetscCall(MatProductSymbolic(adpd)); 1075 1076 adpd_seq = (Mat_SeqAIJ *)((adpd)->data); 1077 adpdi = adpd_seq->i; 1078 adpdj = adpd_seq->j; 1079 p_off = (Mat_SeqAIJ *)((p->B)->data); 1080 poff_i = p_off->i; 1081 poff_j = p_off->j; 1082 1083 /* j_temp stores indices of a result row before they are added to the linked list */ 1084 PetscCall(PetscMalloc1(pN + 2, &j_temp)); 1085 1086 /* Symbolic calc of the A_diag * p_loc_off */ 1087 /* Initial FreeSpace size is fill*(nnz(A)+nnz(P)) */ 1088 PetscCall(PetscFreeSpaceGet(PetscRealIntMultTruncate(fill, PetscIntSumTruncate(adi[am], PetscIntSumTruncate(aoi[am], pi_loc[pm]))), &free_space_diag)); 1089 current_space = free_space_diag; 1090 1091 for (i = 0; i < am; i++) { 1092 /* A_diag * P_loc_off */ 1093 nzi = adi[i + 1] - adi[i]; 1094 for (j = 0; j < nzi; j++) { 1095 row = *adj++; 1096 pnz = poff_i[row + 1] - poff_i[row]; 1097 Jptr = poff_j + poff_i[row]; 1098 for (i1 = 0; i1 < pnz; i1++) j_temp[i1] = p->garray[Jptr[i1]]; 1099 /* add non-zero cols of P into the sorted linked list lnk */ 1100 PetscCall(PetscLLCondensedAddSorted(pnz, j_temp, lnk, lnkbt)); 1101 } 1102 1103 adponz = lnk[0]; 1104 adpoi[i + 1] = adpoi[i] + adponz; 1105 1106 /* if free space is not available, double the total space in the list */ 1107 if (current_space->local_remaining < adponz) { 1108 PetscCall(PetscFreeSpaceGet(PetscIntSumTruncate(adponz, current_space->total_array_size), ¤t_space)); 1109 nspacedouble++; 1110 } 1111 1112 /* Copy data into free space, then initialize lnk */ 1113 PetscCall(PetscLLCondensedClean(pN, adponz, current_space->array, lnk, lnkbt)); 1114 1115 current_space->array += adponz; 1116 current_space->local_used += adponz; 1117 current_space->local_remaining -= adponz; 1118 } 1119 1120 /* Symbolic calc of A_off * P_oth */ 1121 PetscCall(MatSetOptionsPrefix(a->B, prefix)); 1122 PetscCall(MatAppendOptionsPrefix(a->B, "inner_offdiag_")); 1123 PetscCall(MatCreate(PETSC_COMM_SELF, &aopoth)); 1124 PetscCall(MatMatMultSymbolic_SeqAIJ_SeqAIJ(a->B, ptap->P_oth, fill, aopoth)); 1125 aopoth_seq = (Mat_SeqAIJ *)((aopoth)->data); 1126 aopothi = aopoth_seq->i; 1127 aopothj = aopoth_seq->j; 1128 1129 /* Allocate space for apj, adpj, aopj, ... */ 1130 /* destroy lists of free space and other temporary array(s) */ 1131 1132 PetscCall(PetscMalloc1(aopothi[am] + adpoi[am] + adpdi[am] + 2, &ptap->apj)); 1133 PetscCall(PetscMalloc1(adpoi[am] + 2, &adpoj)); 1134 1135 /* Copy from linked list to j-array */ 1136 PetscCall(PetscFreeSpaceContiguous(&free_space_diag, adpoj)); 1137 PetscCall(PetscLLDestroy(lnk, lnkbt)); 1138 1139 adpoJ = adpoj; 1140 adpdJ = adpdj; 1141 aopJ = aopothj; 1142 apj = ptap->apj; 1143 apJ = apj; /* still empty */ 1144 1145 /* Merge j-arrays of A_off * P, A_diag * P_loc_off, and */ 1146 /* A_diag * P_loc_diag to get A*P */ 1147 for (i = 0; i < am; i++) { 1148 aopnz = aopothi[i + 1] - aopothi[i]; 1149 adponz = adpoi[i + 1] - adpoi[i]; 1150 adpdnz = adpdi[i + 1] - adpdi[i]; 1151 1152 /* Correct indices from A_diag*P_diag */ 1153 for (i1 = 0; i1 < adpdnz; i1++) adpdJ[i1] += p_colstart; 1154 /* Merge j-arrays of A_diag * P_loc_off and A_diag * P_loc_diag and A_off * P_oth */ 1155 Merge3SortedArrays(adponz, adpoJ, adpdnz, adpdJ, aopnz, aopJ, &apnz, apJ); 1156 PetscCall(MatPreallocateSet(i + rstart, apnz, apJ, dnz, onz)); 1157 1158 aopJ += aopnz; 1159 adpoJ += adponz; 1160 adpdJ += adpdnz; 1161 apJ += apnz; 1162 api[i + 1] = api[i] + apnz; 1163 } 1164 1165 /* malloc apa to store dense row A[i,:]*P */ 1166 PetscCall(PetscCalloc1(pN + 2, &ptap->apa)); 1167 1168 /* create and assemble symbolic parallel matrix C */ 1169 PetscCall(MatSetSizes(C, am, pn, PETSC_DETERMINE, PETSC_DETERMINE)); 1170 PetscCall(MatSetBlockSizesFromMats(C, A, P)); 1171 PetscCall(MatGetType(A, &mtype)); 1172 PetscCall(MatSetType(C, mtype)); 1173 PetscCall(MatMPIAIJSetPreallocation(C, 0, dnz, 0, onz)); 1174 MatPreallocateEnd(dnz, onz); 1175 1176 PetscCall(MatSetValues_MPIAIJ_CopyFromCSRFormat_Symbolic(C, apj, api)); 1177 PetscCall(MatAssemblyBegin(C, MAT_FINAL_ASSEMBLY)); 1178 PetscCall(MatAssemblyEnd(C, MAT_FINAL_ASSEMBLY)); 1179 PetscCall(MatSetOption(C, MAT_NEW_NONZERO_LOCATION_ERR, PETSC_TRUE)); 1180 1181 C->ops->matmultnumeric = MatMatMultNumeric_MPIAIJ_MPIAIJ_nonscalable; 1182 C->ops->productnumeric = MatProductNumeric_AB; 1183 1184 /* attach the supporting struct to C for reuse */ 1185 C->product->data = ptap; 1186 C->product->destroy = MatDestroy_MPIAIJ_MatMatMult; 1187 1188 /* set MatInfo */ 1189 afill = (PetscReal)api[am] / (adi[am] + aoi[am] + pi_loc[pm] + 1) + 1.e-5; 1190 if (afill < 1.0) afill = 1.0; 1191 C->info.mallocs = nspacedouble; 1192 C->info.fill_ratio_given = fill; 1193 C->info.fill_ratio_needed = afill; 1194 1195 #if defined(PETSC_USE_INFO) 1196 if (api[am]) { 1197 PetscCall(PetscInfo(C, "Reallocs %" PetscInt_FMT "; Fill ratio: given %g needed %g.\n", nspacedouble, (double)fill, (double)afill)); 1198 PetscCall(PetscInfo(C, "Use MatMatMult(A,B,MatReuse,%g,&C) for best performance.;\n", (double)afill)); 1199 } else { 1200 PetscCall(PetscInfo(C, "Empty matrix product\n")); 1201 } 1202 #endif 1203 1204 PetscCall(MatDestroy(&aopoth)); 1205 PetscCall(MatDestroy(&adpd)); 1206 PetscCall(PetscFree(j_temp)); 1207 PetscCall(PetscFree(adpoj)); 1208 PetscCall(PetscFree(adpoi)); 1209 PetscFunctionReturn(PETSC_SUCCESS); 1210 } 1211 1212 /* This routine only works when scall=MAT_REUSE_MATRIX! */ 1213 PetscErrorCode MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ_matmatmult(Mat P, Mat A, Mat C) 1214 { 1215 Mat_APMPI *ptap; 1216 Mat Pt; 1217 1218 PetscFunctionBegin; 1219 MatCheckProduct(C, 3); 1220 ptap = (Mat_APMPI *)C->product->data; 1221 PetscCheck(ptap, PetscObjectComm((PetscObject)C), PETSC_ERR_ARG_WRONGSTATE, "PtAP cannot be computed. Missing data"); 1222 PetscCheck(ptap->Pt, PetscObjectComm((PetscObject)C), PETSC_ERR_ARG_WRONGSTATE, "PtA cannot be reused. Do not call MatProductClear()"); 1223 1224 Pt = ptap->Pt; 1225 PetscCall(MatTransposeSetPrecursor(P, Pt)); 1226 PetscCall(MatTranspose(P, MAT_REUSE_MATRIX, &Pt)); 1227 PetscCall(MatMatMultNumeric_MPIAIJ_MPIAIJ(Pt, A, C)); 1228 PetscFunctionReturn(PETSC_SUCCESS); 1229 } 1230 1231 /* This routine is modified from MatPtAPSymbolic_MPIAIJ_MPIAIJ() */ 1232 PetscErrorCode MatTransposeMatMultSymbolic_MPIAIJ_MPIAIJ_nonscalable(Mat P, Mat A, PetscReal fill, Mat C) 1233 { 1234 Mat_APMPI *ptap; 1235 Mat_MPIAIJ *p = (Mat_MPIAIJ *)P->data; 1236 MPI_Comm comm; 1237 PetscMPIInt size, rank; 1238 PetscFreeSpaceList free_space = NULL, current_space = NULL; 1239 PetscInt pn = P->cmap->n, aN = A->cmap->N, an = A->cmap->n; 1240 PetscInt *lnk, i, k, nsend, rstart; 1241 PetscBT lnkbt; 1242 PetscMPIInt tagi, tagj, *len_si, *len_s, *len_ri, nrecv; 1243 PETSC_UNUSED PetscMPIInt icompleted = 0; 1244 PetscInt **buf_rj, **buf_ri, **buf_ri_k, row, ncols, *cols; 1245 PetscInt len, proc, *dnz, *onz, *owners, nzi; 1246 PetscInt nrows, *buf_s, *buf_si, *buf_si_i, **nextrow, **nextci; 1247 MPI_Request *swaits, *rwaits; 1248 MPI_Status *sstatus, rstatus; 1249 PetscLayout rowmap; 1250 PetscInt *owners_co, *coi, *coj; /* i and j array of (p->B)^T*A*P - used in the communication */ 1251 PetscMPIInt *len_r, *id_r; /* array of length of comm->size, store send/recv matrix values */ 1252 PetscInt *Jptr, *prmap = p->garray, con, j, Crmax; 1253 Mat_SeqAIJ *a_loc, *c_loc, *c_oth; 1254 PetscHMapI ta; 1255 MatType mtype; 1256 const char *prefix; 1257 1258 PetscFunctionBegin; 1259 PetscCall(PetscObjectGetComm((PetscObject)A, &comm)); 1260 PetscCallMPI(MPI_Comm_size(comm, &size)); 1261 PetscCallMPI(MPI_Comm_rank(comm, &rank)); 1262 1263 /* create symbolic parallel matrix C */ 1264 PetscCall(MatGetType(A, &mtype)); 1265 PetscCall(MatSetType(C, mtype)); 1266 1267 C->ops->transposematmultnumeric = MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ_nonscalable; 1268 1269 /* create struct Mat_APMPI and attached it to C later */ 1270 PetscCall(PetscNew(&ptap)); 1271 1272 /* (0) compute Rd = Pd^T, Ro = Po^T */ 1273 PetscCall(MatTranspose(p->A, MAT_INITIAL_MATRIX, &ptap->Rd)); 1274 PetscCall(MatTranspose(p->B, MAT_INITIAL_MATRIX, &ptap->Ro)); 1275 1276 /* (1) compute symbolic A_loc */ 1277 PetscCall(MatMPIAIJGetLocalMat(A, MAT_INITIAL_MATRIX, &ptap->A_loc)); 1278 1279 /* (2-1) compute symbolic C_oth = Ro*A_loc */ 1280 PetscCall(MatGetOptionsPrefix(A, &prefix)); 1281 PetscCall(MatSetOptionsPrefix(ptap->Ro, prefix)); 1282 PetscCall(MatAppendOptionsPrefix(ptap->Ro, "inner_offdiag_")); 1283 PetscCall(MatCreate(PETSC_COMM_SELF, &ptap->C_oth)); 1284 PetscCall(MatMatMultSymbolic_SeqAIJ_SeqAIJ(ptap->Ro, ptap->A_loc, fill, ptap->C_oth)); 1285 1286 /* (3) send coj of C_oth to other processors */ 1287 /* determine row ownership */ 1288 PetscCall(PetscLayoutCreate(comm, &rowmap)); 1289 rowmap->n = pn; 1290 rowmap->bs = 1; 1291 PetscCall(PetscLayoutSetUp(rowmap)); 1292 owners = rowmap->range; 1293 1294 /* determine the number of messages to send, their lengths */ 1295 PetscCall(PetscMalloc4(size, &len_s, size, &len_si, size, &sstatus, size + 2, &owners_co)); 1296 PetscCall(PetscArrayzero(len_s, size)); 1297 PetscCall(PetscArrayzero(len_si, size)); 1298 1299 c_oth = (Mat_SeqAIJ *)ptap->C_oth->data; 1300 coi = c_oth->i; 1301 coj = c_oth->j; 1302 con = ptap->C_oth->rmap->n; 1303 proc = 0; 1304 for (i = 0; i < con; i++) { 1305 while (prmap[i] >= owners[proc + 1]) proc++; 1306 len_si[proc]++; /* num of rows in Co(=Pt*A) to be sent to [proc] */ 1307 len_s[proc] += coi[i + 1] - coi[i]; /* num of nonzeros in Co to be sent to [proc] */ 1308 } 1309 1310 len = 0; /* max length of buf_si[], see (4) */ 1311 owners_co[0] = 0; 1312 nsend = 0; 1313 for (proc = 0; proc < size; proc++) { 1314 owners_co[proc + 1] = owners_co[proc] + len_si[proc]; 1315 if (len_s[proc]) { 1316 nsend++; 1317 len_si[proc] = 2 * (len_si[proc] + 1); /* length of buf_si to be sent to [proc] */ 1318 len += len_si[proc]; 1319 } 1320 } 1321 1322 /* determine the number and length of messages to receive for coi and coj */ 1323 PetscCall(PetscGatherNumberOfMessages(comm, NULL, len_s, &nrecv)); 1324 PetscCall(PetscGatherMessageLengths2(comm, nsend, nrecv, len_s, len_si, &id_r, &len_r, &len_ri)); 1325 1326 /* post the Irecv and Isend of coj */ 1327 PetscCall(PetscCommGetNewTag(comm, &tagj)); 1328 PetscCall(PetscPostIrecvInt(comm, tagj, nrecv, id_r, len_r, &buf_rj, &rwaits)); 1329 PetscCall(PetscMalloc1(nsend + 1, &swaits)); 1330 for (proc = 0, k = 0; proc < size; proc++) { 1331 if (!len_s[proc]) continue; 1332 i = owners_co[proc]; 1333 PetscCallMPI(MPI_Isend(coj + coi[i], len_s[proc], MPIU_INT, proc, tagj, comm, swaits + k)); 1334 k++; 1335 } 1336 1337 /* (2-2) compute symbolic C_loc = Rd*A_loc */ 1338 PetscCall(MatSetOptionsPrefix(ptap->Rd, prefix)); 1339 PetscCall(MatAppendOptionsPrefix(ptap->Rd, "inner_diag_")); 1340 PetscCall(MatCreate(PETSC_COMM_SELF, &ptap->C_loc)); 1341 PetscCall(MatMatMultSymbolic_SeqAIJ_SeqAIJ(ptap->Rd, ptap->A_loc, fill, ptap->C_loc)); 1342 c_loc = (Mat_SeqAIJ *)ptap->C_loc->data; 1343 1344 /* receives coj are complete */ 1345 for (i = 0; i < nrecv; i++) PetscCallMPI(MPI_Waitany(nrecv, rwaits, &icompleted, &rstatus)); 1346 PetscCall(PetscFree(rwaits)); 1347 if (nsend) PetscCallMPI(MPI_Waitall(nsend, swaits, sstatus)); 1348 1349 /* add received column indices into ta to update Crmax */ 1350 a_loc = (Mat_SeqAIJ *)(ptap->A_loc)->data; 1351 1352 /* create and initialize a linked list */ 1353 PetscCall(PetscHMapICreateWithSize(an, &ta)); /* for compute Crmax */ 1354 MatRowMergeMax_SeqAIJ(a_loc, ptap->A_loc->rmap->N, ta); 1355 1356 for (k = 0; k < nrecv; k++) { /* k-th received message */ 1357 Jptr = buf_rj[k]; 1358 for (j = 0; j < len_r[k]; j++) PetscCall(PetscHMapISet(ta, *(Jptr + j) + 1, 1)); 1359 } 1360 PetscCall(PetscHMapIGetSize(ta, &Crmax)); 1361 PetscCall(PetscHMapIDestroy(&ta)); 1362 1363 /* (4) send and recv coi */ 1364 PetscCall(PetscCommGetNewTag(comm, &tagi)); 1365 PetscCall(PetscPostIrecvInt(comm, tagi, nrecv, id_r, len_ri, &buf_ri, &rwaits)); 1366 PetscCall(PetscMalloc1(len + 1, &buf_s)); 1367 buf_si = buf_s; /* points to the beginning of k-th msg to be sent */ 1368 for (proc = 0, k = 0; proc < size; proc++) { 1369 if (!len_s[proc]) continue; 1370 /* form outgoing message for i-structure: 1371 buf_si[0]: nrows to be sent 1372 [1:nrows]: row index (global) 1373 [nrows+1:2*nrows+1]: i-structure index 1374 */ 1375 nrows = len_si[proc] / 2 - 1; /* num of rows in Co to be sent to [proc] */ 1376 buf_si_i = buf_si + nrows + 1; 1377 buf_si[0] = nrows; 1378 buf_si_i[0] = 0; 1379 nrows = 0; 1380 for (i = owners_co[proc]; i < owners_co[proc + 1]; i++) { 1381 nzi = coi[i + 1] - coi[i]; 1382 buf_si_i[nrows + 1] = buf_si_i[nrows] + nzi; /* i-structure */ 1383 buf_si[nrows + 1] = prmap[i] - owners[proc]; /* local row index */ 1384 nrows++; 1385 } 1386 PetscCallMPI(MPI_Isend(buf_si, len_si[proc], MPIU_INT, proc, tagi, comm, swaits + k)); 1387 k++; 1388 buf_si += len_si[proc]; 1389 } 1390 for (i = 0; i < nrecv; i++) PetscCallMPI(MPI_Waitany(nrecv, rwaits, &icompleted, &rstatus)); 1391 PetscCall(PetscFree(rwaits)); 1392 if (nsend) PetscCallMPI(MPI_Waitall(nsend, swaits, sstatus)); 1393 1394 PetscCall(PetscFree4(len_s, len_si, sstatus, owners_co)); 1395 PetscCall(PetscFree(len_ri)); 1396 PetscCall(PetscFree(swaits)); 1397 PetscCall(PetscFree(buf_s)); 1398 1399 /* (5) compute the local portion of C */ 1400 /* set initial free space to be Crmax, sufficient for holding nonzeros in each row of C */ 1401 PetscCall(PetscFreeSpaceGet(Crmax, &free_space)); 1402 current_space = free_space; 1403 1404 PetscCall(PetscMalloc3(nrecv, &buf_ri_k, nrecv, &nextrow, nrecv, &nextci)); 1405 for (k = 0; k < nrecv; k++) { 1406 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 1407 nrows = *buf_ri_k[k]; 1408 nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */ 1409 nextci[k] = buf_ri_k[k] + (nrows + 1); /* points to the next i-structure of k-th recved i-structure */ 1410 } 1411 1412 MatPreallocateBegin(comm, pn, an, dnz, onz); 1413 PetscCall(PetscLLCondensedCreate(Crmax, aN, &lnk, &lnkbt)); 1414 for (i = 0; i < pn; i++) { /* for each local row of C */ 1415 /* add C_loc into C */ 1416 nzi = c_loc->i[i + 1] - c_loc->i[i]; 1417 Jptr = c_loc->j + c_loc->i[i]; 1418 PetscCall(PetscLLCondensedAddSorted(nzi, Jptr, lnk, lnkbt)); 1419 1420 /* add received col data into lnk */ 1421 for (k = 0; k < nrecv; k++) { /* k-th received message */ 1422 if (i == *nextrow[k]) { /* i-th row */ 1423 nzi = *(nextci[k] + 1) - *nextci[k]; 1424 Jptr = buf_rj[k] + *nextci[k]; 1425 PetscCall(PetscLLCondensedAddSorted(nzi, Jptr, lnk, lnkbt)); 1426 nextrow[k]++; 1427 nextci[k]++; 1428 } 1429 } 1430 1431 /* add missing diagonal entry */ 1432 if (C->force_diagonals) { 1433 k = i + owners[rank]; /* column index */ 1434 PetscCall(PetscLLCondensedAddSorted(1, &k, lnk, lnkbt)); 1435 } 1436 1437 nzi = lnk[0]; 1438 1439 /* copy data into free space, then initialize lnk */ 1440 PetscCall(PetscLLCondensedClean(aN, nzi, current_space->array, lnk, lnkbt)); 1441 PetscCall(MatPreallocateSet(i + owners[rank], nzi, current_space->array, dnz, onz)); 1442 } 1443 PetscCall(PetscFree3(buf_ri_k, nextrow, nextci)); 1444 PetscCall(PetscLLDestroy(lnk, lnkbt)); 1445 PetscCall(PetscFreeSpaceDestroy(free_space)); 1446 1447 /* local sizes and preallocation */ 1448 PetscCall(MatSetSizes(C, pn, an, PETSC_DETERMINE, PETSC_DETERMINE)); 1449 if (P->cmap->bs > 0) PetscCall(PetscLayoutSetBlockSize(C->rmap, P->cmap->bs)); 1450 if (A->cmap->bs > 0) PetscCall(PetscLayoutSetBlockSize(C->cmap, A->cmap->bs)); 1451 PetscCall(MatMPIAIJSetPreallocation(C, 0, dnz, 0, onz)); 1452 MatPreallocateEnd(dnz, onz); 1453 1454 /* add C_loc and C_oth to C */ 1455 PetscCall(MatGetOwnershipRange(C, &rstart, NULL)); 1456 for (i = 0; i < pn; i++) { 1457 ncols = c_loc->i[i + 1] - c_loc->i[i]; 1458 cols = c_loc->j + c_loc->i[i]; 1459 row = rstart + i; 1460 PetscCall(MatSetValues(C, 1, (const PetscInt *)&row, ncols, (const PetscInt *)cols, NULL, INSERT_VALUES)); 1461 1462 if (C->force_diagonals) PetscCall(MatSetValues(C, 1, (const PetscInt *)&row, 1, (const PetscInt *)&row, NULL, INSERT_VALUES)); 1463 } 1464 for (i = 0; i < con; i++) { 1465 ncols = c_oth->i[i + 1] - c_oth->i[i]; 1466 cols = c_oth->j + c_oth->i[i]; 1467 row = prmap[i]; 1468 PetscCall(MatSetValues(C, 1, (const PetscInt *)&row, ncols, (const PetscInt *)cols, NULL, INSERT_VALUES)); 1469 } 1470 PetscCall(MatAssemblyBegin(C, MAT_FINAL_ASSEMBLY)); 1471 PetscCall(MatAssemblyEnd(C, MAT_FINAL_ASSEMBLY)); 1472 PetscCall(MatSetOption(C, MAT_NEW_NONZERO_LOCATION_ERR, PETSC_TRUE)); 1473 1474 /* members in merge */ 1475 PetscCall(PetscFree(id_r)); 1476 PetscCall(PetscFree(len_r)); 1477 PetscCall(PetscFree(buf_ri[0])); 1478 PetscCall(PetscFree(buf_ri)); 1479 PetscCall(PetscFree(buf_rj[0])); 1480 PetscCall(PetscFree(buf_rj)); 1481 PetscCall(PetscLayoutDestroy(&rowmap)); 1482 1483 /* attach the supporting struct to C for reuse */ 1484 C->product->data = ptap; 1485 C->product->destroy = MatDestroy_MPIAIJ_PtAP; 1486 PetscFunctionReturn(PETSC_SUCCESS); 1487 } 1488 1489 PetscErrorCode MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ_nonscalable(Mat P, Mat A, Mat C) 1490 { 1491 Mat_MPIAIJ *p = (Mat_MPIAIJ *)P->data; 1492 Mat_SeqAIJ *c_seq; 1493 Mat_APMPI *ptap; 1494 Mat A_loc, C_loc, C_oth; 1495 PetscInt i, rstart, rend, cm, ncols, row; 1496 const PetscInt *cols; 1497 const PetscScalar *vals; 1498 1499 PetscFunctionBegin; 1500 MatCheckProduct(C, 3); 1501 ptap = (Mat_APMPI *)C->product->data; 1502 PetscCheck(ptap, PetscObjectComm((PetscObject)C), PETSC_ERR_ARG_WRONGSTATE, "PtAP cannot be computed. Missing data"); 1503 PetscCheck(ptap->A_loc, PetscObjectComm((PetscObject)C), PETSC_ERR_ARG_WRONGSTATE, "PtA cannot be reused. Do not call MatProductClear()"); 1504 PetscCall(MatZeroEntries(C)); 1505 1506 /* These matrices are obtained in MatTransposeMatMultSymbolic() */ 1507 /* 1) get R = Pd^T, Ro = Po^T */ 1508 PetscCall(MatTransposeSetPrecursor(p->A, ptap->Rd)); 1509 PetscCall(MatTranspose(p->A, MAT_REUSE_MATRIX, &ptap->Rd)); 1510 PetscCall(MatTransposeSetPrecursor(p->B, ptap->Ro)); 1511 PetscCall(MatTranspose(p->B, MAT_REUSE_MATRIX, &ptap->Ro)); 1512 1513 /* 2) compute numeric A_loc */ 1514 PetscCall(MatMPIAIJGetLocalMat(A, MAT_REUSE_MATRIX, &ptap->A_loc)); 1515 1516 /* 3) C_loc = Rd*A_loc, C_oth = Ro*A_loc */ 1517 A_loc = ptap->A_loc; 1518 PetscCall(((ptap->C_loc)->ops->matmultnumeric)(ptap->Rd, A_loc, ptap->C_loc)); 1519 PetscCall(((ptap->C_oth)->ops->matmultnumeric)(ptap->Ro, A_loc, ptap->C_oth)); 1520 C_loc = ptap->C_loc; 1521 C_oth = ptap->C_oth; 1522 1523 /* add C_loc and C_oth to C */ 1524 PetscCall(MatGetOwnershipRange(C, &rstart, &rend)); 1525 1526 /* C_loc -> C */ 1527 cm = C_loc->rmap->N; 1528 c_seq = (Mat_SeqAIJ *)C_loc->data; 1529 cols = c_seq->j; 1530 vals = c_seq->a; 1531 for (i = 0; i < cm; i++) { 1532 ncols = c_seq->i[i + 1] - c_seq->i[i]; 1533 row = rstart + i; 1534 PetscCall(MatSetValues(C, 1, &row, ncols, cols, vals, ADD_VALUES)); 1535 cols += ncols; 1536 vals += ncols; 1537 } 1538 1539 /* Co -> C, off-processor part */ 1540 cm = C_oth->rmap->N; 1541 c_seq = (Mat_SeqAIJ *)C_oth->data; 1542 cols = c_seq->j; 1543 vals = c_seq->a; 1544 for (i = 0; i < cm; i++) { 1545 ncols = c_seq->i[i + 1] - c_seq->i[i]; 1546 row = p->garray[i]; 1547 PetscCall(MatSetValues(C, 1, &row, ncols, cols, vals, ADD_VALUES)); 1548 cols += ncols; 1549 vals += ncols; 1550 } 1551 PetscCall(MatAssemblyBegin(C, MAT_FINAL_ASSEMBLY)); 1552 PetscCall(MatAssemblyEnd(C, MAT_FINAL_ASSEMBLY)); 1553 PetscCall(MatSetOption(C, MAT_NEW_NONZERO_LOCATION_ERR, PETSC_TRUE)); 1554 PetscFunctionReturn(PETSC_SUCCESS); 1555 } 1556 1557 PetscErrorCode MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ(Mat P, Mat A, Mat C) 1558 { 1559 Mat_Merge_SeqsToMPI *merge; 1560 Mat_MPIAIJ *p = (Mat_MPIAIJ *)P->data; 1561 Mat_SeqAIJ *pd = (Mat_SeqAIJ *)(p->A)->data, *po = (Mat_SeqAIJ *)(p->B)->data; 1562 Mat_APMPI *ap; 1563 PetscInt *adj; 1564 PetscInt i, j, k, anz, pnz, row, *cj, nexta; 1565 MatScalar *ada, *ca, valtmp; 1566 PetscInt am = A->rmap->n, cm = C->rmap->n, pon = (p->B)->cmap->n; 1567 MPI_Comm comm; 1568 PetscMPIInt size, rank, taga, *len_s; 1569 PetscInt *owners, proc, nrows, **buf_ri_k, **nextrow, **nextci; 1570 PetscInt **buf_ri, **buf_rj; 1571 PetscInt cnz = 0, *bj_i, *bi, *bj, bnz, nextcj; /* bi,bj,ba: local array of C(mpi mat) */ 1572 MPI_Request *s_waits, *r_waits; 1573 MPI_Status *status; 1574 MatScalar **abuf_r, *ba_i, *pA, *coa, *ba; 1575 const PetscScalar *dummy; 1576 PetscInt *ai, *aj, *coi, *coj, *poJ, *pdJ; 1577 Mat A_loc; 1578 Mat_SeqAIJ *a_loc; 1579 1580 PetscFunctionBegin; 1581 MatCheckProduct(C, 3); 1582 ap = (Mat_APMPI *)C->product->data; 1583 PetscCheck(ap, PetscObjectComm((PetscObject)C), PETSC_ERR_ARG_WRONGSTATE, "PtA cannot be computed. Missing data"); 1584 PetscCheck(ap->A_loc, PetscObjectComm((PetscObject)C), PETSC_ERR_ARG_WRONGSTATE, "PtA cannot be reused. Do not call MatProductClear()"); 1585 PetscCall(PetscObjectGetComm((PetscObject)C, &comm)); 1586 PetscCallMPI(MPI_Comm_size(comm, &size)); 1587 PetscCallMPI(MPI_Comm_rank(comm, &rank)); 1588 1589 merge = ap->merge; 1590 1591 /* 2) compute numeric C_seq = P_loc^T*A_loc */ 1592 /* get data from symbolic products */ 1593 coi = merge->coi; 1594 coj = merge->coj; 1595 PetscCall(PetscCalloc1(coi[pon] + 1, &coa)); 1596 bi = merge->bi; 1597 bj = merge->bj; 1598 owners = merge->rowmap->range; 1599 PetscCall(PetscCalloc1(bi[cm] + 1, &ba)); 1600 1601 /* get A_loc by taking all local rows of A */ 1602 A_loc = ap->A_loc; 1603 PetscCall(MatMPIAIJGetLocalMat(A, MAT_REUSE_MATRIX, &A_loc)); 1604 a_loc = (Mat_SeqAIJ *)(A_loc)->data; 1605 ai = a_loc->i; 1606 aj = a_loc->j; 1607 1608 /* trigger copy to CPU */ 1609 PetscCall(MatSeqAIJGetArrayRead(p->A, &dummy)); 1610 PetscCall(MatSeqAIJRestoreArrayRead(p->A, &dummy)); 1611 PetscCall(MatSeqAIJGetArrayRead(p->B, &dummy)); 1612 PetscCall(MatSeqAIJRestoreArrayRead(p->B, &dummy)); 1613 for (i = 0; i < am; i++) { 1614 anz = ai[i + 1] - ai[i]; 1615 adj = aj + ai[i]; 1616 ada = a_loc->a + ai[i]; 1617 1618 /* 2-b) Compute Cseq = P_loc[i,:]^T*A[i,:] using outer product */ 1619 /* put the value into Co=(p->B)^T*A (off-diagonal part, send to others) */ 1620 pnz = po->i[i + 1] - po->i[i]; 1621 poJ = po->j + po->i[i]; 1622 pA = po->a + po->i[i]; 1623 for (j = 0; j < pnz; j++) { 1624 row = poJ[j]; 1625 cj = coj + coi[row]; 1626 ca = coa + coi[row]; 1627 /* perform sparse axpy */ 1628 nexta = 0; 1629 valtmp = pA[j]; 1630 for (k = 0; nexta < anz; k++) { 1631 if (cj[k] == adj[nexta]) { 1632 ca[k] += valtmp * ada[nexta]; 1633 nexta++; 1634 } 1635 } 1636 PetscCall(PetscLogFlops(2.0 * anz)); 1637 } 1638 1639 /* put the value into Cd (diagonal part) */ 1640 pnz = pd->i[i + 1] - pd->i[i]; 1641 pdJ = pd->j + pd->i[i]; 1642 pA = pd->a + pd->i[i]; 1643 for (j = 0; j < pnz; j++) { 1644 row = pdJ[j]; 1645 cj = bj + bi[row]; 1646 ca = ba + bi[row]; 1647 /* perform sparse axpy */ 1648 nexta = 0; 1649 valtmp = pA[j]; 1650 for (k = 0; nexta < anz; k++) { 1651 if (cj[k] == adj[nexta]) { 1652 ca[k] += valtmp * ada[nexta]; 1653 nexta++; 1654 } 1655 } 1656 PetscCall(PetscLogFlops(2.0 * anz)); 1657 } 1658 } 1659 1660 /* 3) send and recv matrix values coa */ 1661 buf_ri = merge->buf_ri; 1662 buf_rj = merge->buf_rj; 1663 len_s = merge->len_s; 1664 PetscCall(PetscCommGetNewTag(comm, &taga)); 1665 PetscCall(PetscPostIrecvScalar(comm, taga, merge->nrecv, merge->id_r, merge->len_r, &abuf_r, &r_waits)); 1666 1667 PetscCall(PetscMalloc2(merge->nsend + 1, &s_waits, size, &status)); 1668 for (proc = 0, k = 0; proc < size; proc++) { 1669 if (!len_s[proc]) continue; 1670 i = merge->owners_co[proc]; 1671 PetscCallMPI(MPI_Isend(coa + coi[i], len_s[proc], MPIU_MATSCALAR, proc, taga, comm, s_waits + k)); 1672 k++; 1673 } 1674 if (merge->nrecv) PetscCallMPI(MPI_Waitall(merge->nrecv, r_waits, status)); 1675 if (merge->nsend) PetscCallMPI(MPI_Waitall(merge->nsend, s_waits, status)); 1676 1677 PetscCall(PetscFree2(s_waits, status)); 1678 PetscCall(PetscFree(r_waits)); 1679 PetscCall(PetscFree(coa)); 1680 1681 /* 4) insert local Cseq and received values into Cmpi */ 1682 PetscCall(PetscMalloc3(merge->nrecv, &buf_ri_k, merge->nrecv, &nextrow, merge->nrecv, &nextci)); 1683 for (k = 0; k < merge->nrecv; k++) { 1684 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 1685 nrows = *(buf_ri_k[k]); 1686 nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */ 1687 nextci[k] = buf_ri_k[k] + (nrows + 1); /* points to the next i-structure of k-th recved i-structure */ 1688 } 1689 1690 for (i = 0; i < cm; i++) { 1691 row = owners[rank] + i; /* global row index of C_seq */ 1692 bj_i = bj + bi[i]; /* col indices of the i-th row of C */ 1693 ba_i = ba + bi[i]; 1694 bnz = bi[i + 1] - bi[i]; 1695 /* add received vals into ba */ 1696 for (k = 0; k < merge->nrecv; k++) { /* k-th received message */ 1697 /* i-th row */ 1698 if (i == *nextrow[k]) { 1699 cnz = *(nextci[k] + 1) - *nextci[k]; 1700 cj = buf_rj[k] + *(nextci[k]); 1701 ca = abuf_r[k] + *(nextci[k]); 1702 nextcj = 0; 1703 for (j = 0; nextcj < cnz; j++) { 1704 if (bj_i[j] == cj[nextcj]) { /* bcol == ccol */ 1705 ba_i[j] += ca[nextcj++]; 1706 } 1707 } 1708 nextrow[k]++; 1709 nextci[k]++; 1710 PetscCall(PetscLogFlops(2.0 * cnz)); 1711 } 1712 } 1713 PetscCall(MatSetValues(C, 1, &row, bnz, bj_i, ba_i, INSERT_VALUES)); 1714 } 1715 PetscCall(MatAssemblyBegin(C, MAT_FINAL_ASSEMBLY)); 1716 PetscCall(MatAssemblyEnd(C, MAT_FINAL_ASSEMBLY)); 1717 1718 PetscCall(PetscFree(ba)); 1719 PetscCall(PetscFree(abuf_r[0])); 1720 PetscCall(PetscFree(abuf_r)); 1721 PetscCall(PetscFree3(buf_ri_k, nextrow, nextci)); 1722 PetscFunctionReturn(PETSC_SUCCESS); 1723 } 1724 1725 PetscErrorCode MatTransposeMatMultSymbolic_MPIAIJ_MPIAIJ(Mat P, Mat A, PetscReal fill, Mat C) 1726 { 1727 Mat A_loc; 1728 Mat_APMPI *ap; 1729 PetscFreeSpaceList free_space = NULL, current_space = NULL; 1730 Mat_MPIAIJ *p = (Mat_MPIAIJ *)P->data, *a = (Mat_MPIAIJ *)A->data; 1731 PetscInt *pdti, *pdtj, *poti, *potj, *ptJ; 1732 PetscInt nnz; 1733 PetscInt *lnk, *owners_co, *coi, *coj, i, k, pnz, row; 1734 PetscInt am = A->rmap->n, pn = P->cmap->n; 1735 MPI_Comm comm; 1736 PetscMPIInt size, rank, tagi, tagj, *len_si, *len_s, *len_ri; 1737 PetscInt **buf_rj, **buf_ri, **buf_ri_k; 1738 PetscInt len, proc, *dnz, *onz, *owners; 1739 PetscInt nzi, *bi, *bj; 1740 PetscInt nrows, *buf_s, *buf_si, *buf_si_i, **nextrow, **nextci; 1741 MPI_Request *swaits, *rwaits; 1742 MPI_Status *sstatus, rstatus; 1743 Mat_Merge_SeqsToMPI *merge; 1744 PetscInt *ai, *aj, *Jptr, anz, *prmap = p->garray, pon, nspacedouble = 0, j; 1745 PetscReal afill = 1.0, afill_tmp; 1746 PetscInt rstart = P->cmap->rstart, rmax, Armax; 1747 Mat_SeqAIJ *a_loc; 1748 PetscHMapI ta; 1749 MatType mtype; 1750 1751 PetscFunctionBegin; 1752 PetscCall(PetscObjectGetComm((PetscObject)A, &comm)); 1753 /* check if matrix local sizes are compatible */ 1754 PetscCheck(A->rmap->rstart == P->rmap->rstart && A->rmap->rend == P->rmap->rend, comm, PETSC_ERR_ARG_SIZ, "Matrix local dimensions are incompatible, A (%" PetscInt_FMT ", %" PetscInt_FMT ") != P (%" PetscInt_FMT ",%" PetscInt_FMT ")", A->rmap->rstart, 1755 A->rmap->rend, P->rmap->rstart, P->rmap->rend); 1756 1757 PetscCallMPI(MPI_Comm_size(comm, &size)); 1758 PetscCallMPI(MPI_Comm_rank(comm, &rank)); 1759 1760 /* create struct Mat_APMPI and attached it to C later */ 1761 PetscCall(PetscNew(&ap)); 1762 1763 /* get A_loc by taking all local rows of A */ 1764 PetscCall(MatMPIAIJGetLocalMat(A, MAT_INITIAL_MATRIX, &A_loc)); 1765 1766 ap->A_loc = A_loc; 1767 a_loc = (Mat_SeqAIJ *)(A_loc)->data; 1768 ai = a_loc->i; 1769 aj = a_loc->j; 1770 1771 /* determine symbolic Co=(p->B)^T*A - send to others */ 1772 PetscCall(MatGetSymbolicTranspose_SeqAIJ(p->A, &pdti, &pdtj)); 1773 PetscCall(MatGetSymbolicTranspose_SeqAIJ(p->B, &poti, &potj)); 1774 pon = (p->B)->cmap->n; /* total num of rows to be sent to other processors 1775 >= (num of nonzero rows of C_seq) - pn */ 1776 PetscCall(PetscMalloc1(pon + 1, &coi)); 1777 coi[0] = 0; 1778 1779 /* set initial free space to be fill*(nnz(p->B) + nnz(A)) */ 1780 nnz = PetscRealIntMultTruncate(fill, PetscIntSumTruncate(poti[pon], ai[am])); 1781 PetscCall(PetscFreeSpaceGet(nnz, &free_space)); 1782 current_space = free_space; 1783 1784 /* create and initialize a linked list */ 1785 PetscCall(PetscHMapICreateWithSize(A->cmap->n + a->B->cmap->N, &ta)); 1786 MatRowMergeMax_SeqAIJ(a_loc, am, ta); 1787 PetscCall(PetscHMapIGetSize(ta, &Armax)); 1788 1789 PetscCall(PetscLLCondensedCreate_Scalable(Armax, &lnk)); 1790 1791 for (i = 0; i < pon; i++) { 1792 pnz = poti[i + 1] - poti[i]; 1793 ptJ = potj + poti[i]; 1794 for (j = 0; j < pnz; j++) { 1795 row = ptJ[j]; /* row of A_loc == col of Pot */ 1796 anz = ai[row + 1] - ai[row]; 1797 Jptr = aj + ai[row]; 1798 /* add non-zero cols of AP into the sorted linked list lnk */ 1799 PetscCall(PetscLLCondensedAddSorted_Scalable(anz, Jptr, lnk)); 1800 } 1801 nnz = lnk[0]; 1802 1803 /* If free space is not available, double the total space in the list */ 1804 if (current_space->local_remaining < nnz) { 1805 PetscCall(PetscFreeSpaceGet(PetscIntSumTruncate(nnz, current_space->total_array_size), ¤t_space)); 1806 nspacedouble++; 1807 } 1808 1809 /* Copy data into free space, and zero out denserows */ 1810 PetscCall(PetscLLCondensedClean_Scalable(nnz, current_space->array, lnk)); 1811 1812 current_space->array += nnz; 1813 current_space->local_used += nnz; 1814 current_space->local_remaining -= nnz; 1815 1816 coi[i + 1] = coi[i] + nnz; 1817 } 1818 1819 PetscCall(PetscMalloc1(coi[pon] + 1, &coj)); 1820 PetscCall(PetscFreeSpaceContiguous(&free_space, coj)); 1821 PetscCall(PetscLLCondensedDestroy_Scalable(lnk)); /* must destroy to get a new one for C */ 1822 1823 afill_tmp = (PetscReal)coi[pon] / (poti[pon] + ai[am] + 1); 1824 if (afill_tmp > afill) afill = afill_tmp; 1825 1826 /* send j-array (coj) of Co to other processors */ 1827 /* determine row ownership */ 1828 PetscCall(PetscNew(&merge)); 1829 PetscCall(PetscLayoutCreate(comm, &merge->rowmap)); 1830 1831 merge->rowmap->n = pn; 1832 merge->rowmap->bs = 1; 1833 1834 PetscCall(PetscLayoutSetUp(merge->rowmap)); 1835 owners = merge->rowmap->range; 1836 1837 /* determine the number of messages to send, their lengths */ 1838 PetscCall(PetscCalloc1(size, &len_si)); 1839 PetscCall(PetscCalloc1(size, &merge->len_s)); 1840 1841 len_s = merge->len_s; 1842 merge->nsend = 0; 1843 1844 PetscCall(PetscMalloc1(size + 2, &owners_co)); 1845 1846 proc = 0; 1847 for (i = 0; i < pon; i++) { 1848 while (prmap[i] >= owners[proc + 1]) proc++; 1849 len_si[proc]++; /* num of rows in Co to be sent to [proc] */ 1850 len_s[proc] += coi[i + 1] - coi[i]; 1851 } 1852 1853 len = 0; /* max length of buf_si[] */ 1854 owners_co[0] = 0; 1855 for (proc = 0; proc < size; proc++) { 1856 owners_co[proc + 1] = owners_co[proc] + len_si[proc]; 1857 if (len_si[proc]) { 1858 merge->nsend++; 1859 len_si[proc] = 2 * (len_si[proc] + 1); 1860 len += len_si[proc]; 1861 } 1862 } 1863 1864 /* determine the number and length of messages to receive for coi and coj */ 1865 PetscCall(PetscGatherNumberOfMessages(comm, NULL, len_s, &merge->nrecv)); 1866 PetscCall(PetscGatherMessageLengths2(comm, merge->nsend, merge->nrecv, len_s, len_si, &merge->id_r, &merge->len_r, &len_ri)); 1867 1868 /* post the Irecv and Isend of coj */ 1869 PetscCall(PetscCommGetNewTag(comm, &tagj)); 1870 PetscCall(PetscPostIrecvInt(comm, tagj, merge->nrecv, merge->id_r, merge->len_r, &buf_rj, &rwaits)); 1871 PetscCall(PetscMalloc1(merge->nsend + 1, &swaits)); 1872 for (proc = 0, k = 0; proc < size; proc++) { 1873 if (!len_s[proc]) continue; 1874 i = owners_co[proc]; 1875 PetscCallMPI(MPI_Isend(coj + coi[i], len_s[proc], MPIU_INT, proc, tagj, comm, swaits + k)); 1876 k++; 1877 } 1878 1879 /* receives and sends of coj are complete */ 1880 PetscCall(PetscMalloc1(size, &sstatus)); 1881 for (i = 0; i < merge->nrecv; i++) { 1882 PETSC_UNUSED PetscMPIInt icompleted; 1883 PetscCallMPI(MPI_Waitany(merge->nrecv, rwaits, &icompleted, &rstatus)); 1884 } 1885 PetscCall(PetscFree(rwaits)); 1886 if (merge->nsend) PetscCallMPI(MPI_Waitall(merge->nsend, swaits, sstatus)); 1887 1888 /* add received column indices into table to update Armax */ 1889 /* Armax can be as large as aN if a P[row,:] is dense, see src/ksp/ksp/tutorials/ex56.c! */ 1890 for (k = 0; k < merge->nrecv; k++) { /* k-th received message */ 1891 Jptr = buf_rj[k]; 1892 for (j = 0; j < merge->len_r[k]; j++) PetscCall(PetscHMapISet(ta, *(Jptr + j) + 1, 1)); 1893 } 1894 PetscCall(PetscHMapIGetSize(ta, &Armax)); 1895 1896 /* send and recv coi */ 1897 PetscCall(PetscCommGetNewTag(comm, &tagi)); 1898 PetscCall(PetscPostIrecvInt(comm, tagi, merge->nrecv, merge->id_r, len_ri, &buf_ri, &rwaits)); 1899 PetscCall(PetscMalloc1(len + 1, &buf_s)); 1900 buf_si = buf_s; /* points to the beginning of k-th msg to be sent */ 1901 for (proc = 0, k = 0; proc < size; proc++) { 1902 if (!len_s[proc]) continue; 1903 /* form outgoing message for i-structure: 1904 buf_si[0]: nrows to be sent 1905 [1:nrows]: row index (global) 1906 [nrows+1:2*nrows+1]: i-structure index 1907 */ 1908 nrows = len_si[proc] / 2 - 1; 1909 buf_si_i = buf_si + nrows + 1; 1910 buf_si[0] = nrows; 1911 buf_si_i[0] = 0; 1912 nrows = 0; 1913 for (i = owners_co[proc]; i < owners_co[proc + 1]; i++) { 1914 nzi = coi[i + 1] - coi[i]; 1915 buf_si_i[nrows + 1] = buf_si_i[nrows] + nzi; /* i-structure */ 1916 buf_si[nrows + 1] = prmap[i] - owners[proc]; /* local row index */ 1917 nrows++; 1918 } 1919 PetscCallMPI(MPI_Isend(buf_si, len_si[proc], MPIU_INT, proc, tagi, comm, swaits + k)); 1920 k++; 1921 buf_si += len_si[proc]; 1922 } 1923 i = merge->nrecv; 1924 while (i--) { 1925 PETSC_UNUSED PetscMPIInt icompleted; 1926 PetscCallMPI(MPI_Waitany(merge->nrecv, rwaits, &icompleted, &rstatus)); 1927 } 1928 PetscCall(PetscFree(rwaits)); 1929 if (merge->nsend) PetscCallMPI(MPI_Waitall(merge->nsend, swaits, sstatus)); 1930 PetscCall(PetscFree(len_si)); 1931 PetscCall(PetscFree(len_ri)); 1932 PetscCall(PetscFree(swaits)); 1933 PetscCall(PetscFree(sstatus)); 1934 PetscCall(PetscFree(buf_s)); 1935 1936 /* compute the local portion of C (mpi mat) */ 1937 /* allocate bi array and free space for accumulating nonzero column info */ 1938 PetscCall(PetscMalloc1(pn + 1, &bi)); 1939 bi[0] = 0; 1940 1941 /* set initial free space to be fill*(nnz(P) + nnz(AP)) */ 1942 nnz = PetscRealIntMultTruncate(fill, PetscIntSumTruncate(pdti[pn], PetscIntSumTruncate(poti[pon], ai[am]))); 1943 PetscCall(PetscFreeSpaceGet(nnz, &free_space)); 1944 current_space = free_space; 1945 1946 PetscCall(PetscMalloc3(merge->nrecv, &buf_ri_k, merge->nrecv, &nextrow, merge->nrecv, &nextci)); 1947 for (k = 0; k < merge->nrecv; k++) { 1948 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 1949 nrows = *buf_ri_k[k]; 1950 nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */ 1951 nextci[k] = buf_ri_k[k] + (nrows + 1); /* points to the next i-structure of k-th received i-structure */ 1952 } 1953 1954 PetscCall(PetscLLCondensedCreate_Scalable(Armax, &lnk)); 1955 MatPreallocateBegin(comm, pn, A->cmap->n, dnz, onz); 1956 rmax = 0; 1957 for (i = 0; i < pn; i++) { 1958 /* add pdt[i,:]*AP into lnk */ 1959 pnz = pdti[i + 1] - pdti[i]; 1960 ptJ = pdtj + pdti[i]; 1961 for (j = 0; j < pnz; j++) { 1962 row = ptJ[j]; /* row of AP == col of Pt */ 1963 anz = ai[row + 1] - ai[row]; 1964 Jptr = aj + ai[row]; 1965 /* add non-zero cols of AP into the sorted linked list lnk */ 1966 PetscCall(PetscLLCondensedAddSorted_Scalable(anz, Jptr, lnk)); 1967 } 1968 1969 /* add received col data into lnk */ 1970 for (k = 0; k < merge->nrecv; k++) { /* k-th received message */ 1971 if (i == *nextrow[k]) { /* i-th row */ 1972 nzi = *(nextci[k] + 1) - *nextci[k]; 1973 Jptr = buf_rj[k] + *nextci[k]; 1974 PetscCall(PetscLLCondensedAddSorted_Scalable(nzi, Jptr, lnk)); 1975 nextrow[k]++; 1976 nextci[k]++; 1977 } 1978 } 1979 1980 /* add missing diagonal entry */ 1981 if (C->force_diagonals) { 1982 k = i + owners[rank]; /* column index */ 1983 PetscCall(PetscLLCondensedAddSorted_Scalable(1, &k, lnk)); 1984 } 1985 1986 nnz = lnk[0]; 1987 1988 /* if free space is not available, make more free space */ 1989 if (current_space->local_remaining < nnz) { 1990 PetscCall(PetscFreeSpaceGet(PetscIntSumTruncate(nnz, current_space->total_array_size), ¤t_space)); 1991 nspacedouble++; 1992 } 1993 /* copy data into free space, then initialize lnk */ 1994 PetscCall(PetscLLCondensedClean_Scalable(nnz, current_space->array, lnk)); 1995 PetscCall(MatPreallocateSet(i + owners[rank], nnz, current_space->array, dnz, onz)); 1996 1997 current_space->array += nnz; 1998 current_space->local_used += nnz; 1999 current_space->local_remaining -= nnz; 2000 2001 bi[i + 1] = bi[i] + nnz; 2002 if (nnz > rmax) rmax = nnz; 2003 } 2004 PetscCall(PetscFree3(buf_ri_k, nextrow, nextci)); 2005 2006 PetscCall(PetscMalloc1(bi[pn] + 1, &bj)); 2007 PetscCall(PetscFreeSpaceContiguous(&free_space, bj)); 2008 afill_tmp = (PetscReal)bi[pn] / (pdti[pn] + poti[pon] + ai[am] + 1); 2009 if (afill_tmp > afill) afill = afill_tmp; 2010 PetscCall(PetscLLCondensedDestroy_Scalable(lnk)); 2011 PetscCall(PetscHMapIDestroy(&ta)); 2012 PetscCall(MatRestoreSymbolicTranspose_SeqAIJ(p->A, &pdti, &pdtj)); 2013 PetscCall(MatRestoreSymbolicTranspose_SeqAIJ(p->B, &poti, &potj)); 2014 2015 /* create symbolic parallel matrix C - why cannot be assembled in Numeric part */ 2016 PetscCall(MatSetSizes(C, pn, A->cmap->n, PETSC_DETERMINE, PETSC_DETERMINE)); 2017 PetscCall(MatSetBlockSizes(C, PetscAbs(P->cmap->bs), PetscAbs(A->cmap->bs))); 2018 PetscCall(MatGetType(A, &mtype)); 2019 PetscCall(MatSetType(C, mtype)); 2020 PetscCall(MatMPIAIJSetPreallocation(C, 0, dnz, 0, onz)); 2021 MatPreallocateEnd(dnz, onz); 2022 PetscCall(MatSetBlockSize(C, 1)); 2023 PetscCall(MatSetOption(C, MAT_NO_OFF_PROC_ENTRIES, PETSC_TRUE)); 2024 for (i = 0; i < pn; i++) { 2025 row = i + rstart; 2026 nnz = bi[i + 1] - bi[i]; 2027 Jptr = bj + bi[i]; 2028 PetscCall(MatSetValues(C, 1, &row, nnz, Jptr, NULL, INSERT_VALUES)); 2029 } 2030 PetscCall(MatAssemblyBegin(C, MAT_FINAL_ASSEMBLY)); 2031 PetscCall(MatAssemblyEnd(C, MAT_FINAL_ASSEMBLY)); 2032 PetscCall(MatSetOption(C, MAT_NEW_NONZERO_LOCATION_ERR, PETSC_TRUE)); 2033 merge->bi = bi; 2034 merge->bj = bj; 2035 merge->coi = coi; 2036 merge->coj = coj; 2037 merge->buf_ri = buf_ri; 2038 merge->buf_rj = buf_rj; 2039 merge->owners_co = owners_co; 2040 2041 /* attach the supporting struct to C for reuse */ 2042 C->product->data = ap; 2043 C->product->destroy = MatDestroy_MPIAIJ_PtAP; 2044 ap->merge = merge; 2045 2046 C->ops->mattransposemultnumeric = MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ; 2047 2048 #if defined(PETSC_USE_INFO) 2049 if (bi[pn] != 0) { 2050 PetscCall(PetscInfo(C, "Reallocs %" PetscInt_FMT "; Fill ratio: given %g needed %g.\n", nspacedouble, (double)fill, (double)afill)); 2051 PetscCall(PetscInfo(C, "Use MatTransposeMatMult(A,B,MatReuse,%g,&C) for best performance.\n", (double)afill)); 2052 } else { 2053 PetscCall(PetscInfo(C, "Empty matrix product\n")); 2054 } 2055 #endif 2056 PetscFunctionReturn(PETSC_SUCCESS); 2057 } 2058 2059 static PetscErrorCode MatProductSymbolic_AtB_MPIAIJ_MPIAIJ(Mat C) 2060 { 2061 Mat_Product *product = C->product; 2062 Mat A = product->A, B = product->B; 2063 PetscReal fill = product->fill; 2064 PetscBool flg; 2065 2066 PetscFunctionBegin; 2067 /* scalable */ 2068 PetscCall(PetscStrcmp(product->alg, "scalable", &flg)); 2069 if (flg) { 2070 PetscCall(MatTransposeMatMultSymbolic_MPIAIJ_MPIAIJ(A, B, fill, C)); 2071 goto next; 2072 } 2073 2074 /* nonscalable */ 2075 PetscCall(PetscStrcmp(product->alg, "nonscalable", &flg)); 2076 if (flg) { 2077 PetscCall(MatTransposeMatMultSymbolic_MPIAIJ_MPIAIJ_nonscalable(A, B, fill, C)); 2078 goto next; 2079 } 2080 2081 /* matmatmult */ 2082 PetscCall(PetscStrcmp(product->alg, "at*b", &flg)); 2083 if (flg) { 2084 Mat At; 2085 Mat_APMPI *ptap; 2086 2087 PetscCall(MatTranspose(A, MAT_INITIAL_MATRIX, &At)); 2088 PetscCall(MatMatMultSymbolic_MPIAIJ_MPIAIJ(At, B, fill, C)); 2089 ptap = (Mat_APMPI *)C->product->data; 2090 if (ptap) { 2091 ptap->Pt = At; 2092 C->product->destroy = MatDestroy_MPIAIJ_PtAP; 2093 } 2094 C->ops->transposematmultnumeric = MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ_matmatmult; 2095 goto next; 2096 } 2097 2098 /* backend general code */ 2099 PetscCall(PetscStrcmp(product->alg, "backend", &flg)); 2100 if (flg) { 2101 PetscCall(MatProductSymbolic_MPIAIJBACKEND(C)); 2102 PetscFunctionReturn(PETSC_SUCCESS); 2103 } 2104 2105 SETERRQ(PETSC_COMM_SELF, PETSC_ERR_SUP, "MatProduct type is not supported"); 2106 2107 next: 2108 C->ops->productnumeric = MatProductNumeric_AtB; 2109 PetscFunctionReturn(PETSC_SUCCESS); 2110 } 2111 2112 /* Set options for MatMatMultxxx_MPIAIJ_MPIAIJ */ 2113 static PetscErrorCode MatProductSetFromOptions_MPIAIJ_AB(Mat C) 2114 { 2115 Mat_Product *product = C->product; 2116 Mat A = product->A, B = product->B; 2117 #if defined(PETSC_HAVE_HYPRE) 2118 const char *algTypes[5] = {"scalable", "nonscalable", "seqmpi", "backend", "hypre"}; 2119 PetscInt nalg = 5; 2120 #else 2121 const char *algTypes[4] = { 2122 "scalable", 2123 "nonscalable", 2124 "seqmpi", 2125 "backend", 2126 }; 2127 PetscInt nalg = 4; 2128 #endif 2129 PetscInt alg = 1; /* set nonscalable algorithm as default */ 2130 PetscBool flg; 2131 MPI_Comm comm; 2132 2133 PetscFunctionBegin; 2134 PetscCall(PetscObjectGetComm((PetscObject)C, &comm)); 2135 2136 /* Set "nonscalable" as default algorithm */ 2137 PetscCall(PetscStrcmp(C->product->alg, "default", &flg)); 2138 if (flg) { 2139 PetscCall(MatProductSetAlgorithm(C, (MatProductAlgorithm)algTypes[alg])); 2140 2141 /* Set "scalable" as default if BN and local nonzeros of A and B are large */ 2142 if (B->cmap->N > 100000) { /* may switch to scalable algorithm as default */ 2143 MatInfo Ainfo, Binfo; 2144 PetscInt nz_local; 2145 PetscBool alg_scalable_loc = PETSC_FALSE, alg_scalable; 2146 2147 PetscCall(MatGetInfo(A, MAT_LOCAL, &Ainfo)); 2148 PetscCall(MatGetInfo(B, MAT_LOCAL, &Binfo)); 2149 nz_local = (PetscInt)(Ainfo.nz_allocated + Binfo.nz_allocated); 2150 2151 if (B->cmap->N > product->fill * nz_local) alg_scalable_loc = PETSC_TRUE; 2152 PetscCall(MPIU_Allreduce(&alg_scalable_loc, &alg_scalable, 1, MPIU_BOOL, MPI_LOR, comm)); 2153 2154 if (alg_scalable) { 2155 alg = 0; /* scalable algorithm would 50% slower than nonscalable algorithm */ 2156 PetscCall(MatProductSetAlgorithm(C, (MatProductAlgorithm)algTypes[alg])); 2157 PetscCall(PetscInfo(B, "Use scalable algorithm, BN %" PetscInt_FMT ", fill*nz_allocated %g\n", B->cmap->N, (double)(product->fill * nz_local))); 2158 } 2159 } 2160 } 2161 2162 /* Get runtime option */ 2163 if (product->api_user) { 2164 PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatMatMult", "Mat"); 2165 PetscCall(PetscOptionsEList("-matmatmult_via", "Algorithmic approach", "MatMatMult", algTypes, nalg, algTypes[alg], &alg, &flg)); 2166 PetscOptionsEnd(); 2167 } else { 2168 PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatProduct_AB", "Mat"); 2169 PetscCall(PetscOptionsEList("-mat_product_algorithm", "Algorithmic approach", "MatMatMult", algTypes, nalg, algTypes[alg], &alg, &flg)); 2170 PetscOptionsEnd(); 2171 } 2172 if (flg) PetscCall(MatProductSetAlgorithm(C, (MatProductAlgorithm)algTypes[alg])); 2173 2174 C->ops->productsymbolic = MatProductSymbolic_AB_MPIAIJ_MPIAIJ; 2175 PetscFunctionReturn(PETSC_SUCCESS); 2176 } 2177 2178 static PetscErrorCode MatProductSetFromOptions_MPIAIJ_ABt(Mat C) 2179 { 2180 PetscFunctionBegin; 2181 PetscCall(MatProductSetFromOptions_MPIAIJ_AB(C)); 2182 C->ops->productsymbolic = MatProductSymbolic_ABt_MPIAIJ_MPIAIJ; 2183 PetscFunctionReturn(PETSC_SUCCESS); 2184 } 2185 2186 /* Set options for MatTransposeMatMultXXX_MPIAIJ_MPIAIJ */ 2187 static PetscErrorCode MatProductSetFromOptions_MPIAIJ_AtB(Mat C) 2188 { 2189 Mat_Product *product = C->product; 2190 Mat A = product->A, B = product->B; 2191 const char *algTypes[4] = {"scalable", "nonscalable", "at*b", "backend"}; 2192 PetscInt nalg = 4; 2193 PetscInt alg = 1; /* set default algorithm */ 2194 PetscBool flg; 2195 MPI_Comm comm; 2196 2197 PetscFunctionBegin; 2198 /* Check matrix local sizes */ 2199 PetscCall(PetscObjectGetComm((PetscObject)C, &comm)); 2200 PetscCheck(A->rmap->rstart == B->rmap->rstart && A->rmap->rend == B->rmap->rend, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Matrix local dimensions are incompatible, A (%" PetscInt_FMT ", %" PetscInt_FMT ") != B (%" PetscInt_FMT ",%" PetscInt_FMT ")", 2201 A->rmap->rstart, A->rmap->rend, B->rmap->rstart, B->rmap->rend); 2202 2203 /* Set default algorithm */ 2204 PetscCall(PetscStrcmp(C->product->alg, "default", &flg)); 2205 if (flg) PetscCall(MatProductSetAlgorithm(C, (MatProductAlgorithm)algTypes[alg])); 2206 2207 /* Set "scalable" as default if BN and local nonzeros of A and B are large */ 2208 if (alg && B->cmap->N > 100000) { /* may switch to scalable algorithm as default */ 2209 MatInfo Ainfo, Binfo; 2210 PetscInt nz_local; 2211 PetscBool alg_scalable_loc = PETSC_FALSE, alg_scalable; 2212 2213 PetscCall(MatGetInfo(A, MAT_LOCAL, &Ainfo)); 2214 PetscCall(MatGetInfo(B, MAT_LOCAL, &Binfo)); 2215 nz_local = (PetscInt)(Ainfo.nz_allocated + Binfo.nz_allocated); 2216 2217 if (B->cmap->N > product->fill * nz_local) alg_scalable_loc = PETSC_TRUE; 2218 PetscCall(MPIU_Allreduce(&alg_scalable_loc, &alg_scalable, 1, MPIU_BOOL, MPI_LOR, comm)); 2219 2220 if (alg_scalable) { 2221 alg = 0; /* scalable algorithm would 50% slower than nonscalable algorithm */ 2222 PetscCall(MatProductSetAlgorithm(C, (MatProductAlgorithm)algTypes[alg])); 2223 PetscCall(PetscInfo(B, "Use scalable algorithm, BN %" PetscInt_FMT ", fill*nz_allocated %g\n", B->cmap->N, (double)(product->fill * nz_local))); 2224 } 2225 } 2226 2227 /* Get runtime option */ 2228 if (product->api_user) { 2229 PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatTransposeMatMult", "Mat"); 2230 PetscCall(PetscOptionsEList("-mattransposematmult_via", "Algorithmic approach", "MatTransposeMatMult", algTypes, nalg, algTypes[alg], &alg, &flg)); 2231 PetscOptionsEnd(); 2232 } else { 2233 PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatProduct_AtB", "Mat"); 2234 PetscCall(PetscOptionsEList("-mat_product_algorithm", "Algorithmic approach", "MatTransposeMatMult", algTypes, nalg, algTypes[alg], &alg, &flg)); 2235 PetscOptionsEnd(); 2236 } 2237 if (flg) PetscCall(MatProductSetAlgorithm(C, (MatProductAlgorithm)algTypes[alg])); 2238 2239 C->ops->productsymbolic = MatProductSymbolic_AtB_MPIAIJ_MPIAIJ; 2240 PetscFunctionReturn(PETSC_SUCCESS); 2241 } 2242 2243 static PetscErrorCode MatProductSetFromOptions_MPIAIJ_PtAP(Mat C) 2244 { 2245 Mat_Product *product = C->product; 2246 Mat A = product->A, P = product->B; 2247 MPI_Comm comm; 2248 PetscBool flg; 2249 PetscInt alg = 1; /* set default algorithm */ 2250 #if !defined(PETSC_HAVE_HYPRE) 2251 const char *algTypes[5] = {"scalable", "nonscalable", "allatonce", "allatonce_merged", "backend"}; 2252 PetscInt nalg = 5; 2253 #else 2254 const char *algTypes[6] = {"scalable", "nonscalable", "allatonce", "allatonce_merged", "backend", "hypre"}; 2255 PetscInt nalg = 6; 2256 #endif 2257 PetscInt pN = P->cmap->N; 2258 2259 PetscFunctionBegin; 2260 /* Check matrix local sizes */ 2261 PetscCall(PetscObjectGetComm((PetscObject)C, &comm)); 2262 PetscCheck(A->rmap->rstart == P->rmap->rstart && A->rmap->rend == P->rmap->rend, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Matrix local dimensions are incompatible, Arow (%" PetscInt_FMT ", %" PetscInt_FMT ") != Prow (%" PetscInt_FMT ",%" PetscInt_FMT ")", 2263 A->rmap->rstart, A->rmap->rend, P->rmap->rstart, P->rmap->rend); 2264 PetscCheck(A->cmap->rstart == P->rmap->rstart && A->cmap->rend == P->rmap->rend, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Matrix local dimensions are incompatible, Acol (%" PetscInt_FMT ", %" PetscInt_FMT ") != Prow (%" PetscInt_FMT ",%" PetscInt_FMT ")", 2265 A->cmap->rstart, A->cmap->rend, P->rmap->rstart, P->rmap->rend); 2266 2267 /* Set "nonscalable" as default algorithm */ 2268 PetscCall(PetscStrcmp(C->product->alg, "default", &flg)); 2269 if (flg) { 2270 PetscCall(MatProductSetAlgorithm(C, (MatProductAlgorithm)algTypes[alg])); 2271 2272 /* Set "scalable" as default if BN and local nonzeros of A and B are large */ 2273 if (pN > 100000) { 2274 MatInfo Ainfo, Pinfo; 2275 PetscInt nz_local; 2276 PetscBool alg_scalable_loc = PETSC_FALSE, alg_scalable; 2277 2278 PetscCall(MatGetInfo(A, MAT_LOCAL, &Ainfo)); 2279 PetscCall(MatGetInfo(P, MAT_LOCAL, &Pinfo)); 2280 nz_local = (PetscInt)(Ainfo.nz_allocated + Pinfo.nz_allocated); 2281 2282 if (pN > product->fill * nz_local) alg_scalable_loc = PETSC_TRUE; 2283 PetscCall(MPIU_Allreduce(&alg_scalable_loc, &alg_scalable, 1, MPIU_BOOL, MPI_LOR, comm)); 2284 2285 if (alg_scalable) { 2286 alg = 0; /* scalable algorithm would 50% slower than nonscalable algorithm */ 2287 PetscCall(MatProductSetAlgorithm(C, (MatProductAlgorithm)algTypes[alg])); 2288 } 2289 } 2290 } 2291 2292 /* Get runtime option */ 2293 if (product->api_user) { 2294 PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatPtAP", "Mat"); 2295 PetscCall(PetscOptionsEList("-matptap_via", "Algorithmic approach", "MatPtAP", algTypes, nalg, algTypes[alg], &alg, &flg)); 2296 PetscOptionsEnd(); 2297 } else { 2298 PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatProduct_PtAP", "Mat"); 2299 PetscCall(PetscOptionsEList("-mat_product_algorithm", "Algorithmic approach", "MatPtAP", algTypes, nalg, algTypes[alg], &alg, &flg)); 2300 PetscOptionsEnd(); 2301 } 2302 if (flg) PetscCall(MatProductSetAlgorithm(C, (MatProductAlgorithm)algTypes[alg])); 2303 2304 C->ops->productsymbolic = MatProductSymbolic_PtAP_MPIAIJ_MPIAIJ; 2305 PetscFunctionReturn(PETSC_SUCCESS); 2306 } 2307 2308 static PetscErrorCode MatProductSetFromOptions_MPIAIJ_RARt(Mat C) 2309 { 2310 Mat_Product *product = C->product; 2311 Mat A = product->A, R = product->B; 2312 2313 PetscFunctionBegin; 2314 /* Check matrix local sizes */ 2315 PetscCheck(A->cmap->n == R->cmap->n && A->rmap->n == R->cmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Matrix local dimensions are incompatible, A local (%" PetscInt_FMT ", %" PetscInt_FMT "), R local (%" PetscInt_FMT ",%" PetscInt_FMT ")", A->rmap->n, 2316 A->rmap->n, R->rmap->n, R->cmap->n); 2317 2318 C->ops->productsymbolic = MatProductSymbolic_RARt_MPIAIJ_MPIAIJ; 2319 PetscFunctionReturn(PETSC_SUCCESS); 2320 } 2321 2322 /* 2323 Set options for ABC = A*B*C = A*(B*C); ABC's algorithm must be chosen from AB's algorithm 2324 */ 2325 static PetscErrorCode MatProductSetFromOptions_MPIAIJ_ABC(Mat C) 2326 { 2327 Mat_Product *product = C->product; 2328 PetscBool flg = PETSC_FALSE; 2329 PetscInt alg = 1; /* default algorithm */ 2330 const char *algTypes[3] = {"scalable", "nonscalable", "seqmpi"}; 2331 PetscInt nalg = 3; 2332 2333 PetscFunctionBegin; 2334 /* Set default algorithm */ 2335 PetscCall(PetscStrcmp(C->product->alg, "default", &flg)); 2336 if (flg) PetscCall(MatProductSetAlgorithm(C, (MatProductAlgorithm)algTypes[alg])); 2337 2338 /* Get runtime option */ 2339 if (product->api_user) { 2340 PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatMatMatMult", "Mat"); 2341 PetscCall(PetscOptionsEList("-matmatmatmult_via", "Algorithmic approach", "MatMatMatMult", algTypes, nalg, algTypes[alg], &alg, &flg)); 2342 PetscOptionsEnd(); 2343 } else { 2344 PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatProduct_ABC", "Mat"); 2345 PetscCall(PetscOptionsEList("-mat_product_algorithm", "Algorithmic approach", "MatProduct_ABC", algTypes, nalg, algTypes[alg], &alg, &flg)); 2346 PetscOptionsEnd(); 2347 } 2348 if (flg) PetscCall(MatProductSetAlgorithm(C, (MatProductAlgorithm)algTypes[alg])); 2349 2350 C->ops->matmatmultsymbolic = MatMatMatMultSymbolic_MPIAIJ_MPIAIJ_MPIAIJ; 2351 C->ops->productsymbolic = MatProductSymbolic_ABC; 2352 PetscFunctionReturn(PETSC_SUCCESS); 2353 } 2354 2355 PETSC_INTERN PetscErrorCode MatProductSetFromOptions_MPIAIJ(Mat C) 2356 { 2357 Mat_Product *product = C->product; 2358 2359 PetscFunctionBegin; 2360 switch (product->type) { 2361 case MATPRODUCT_AB: 2362 PetscCall(MatProductSetFromOptions_MPIAIJ_AB(C)); 2363 break; 2364 case MATPRODUCT_ABt: 2365 PetscCall(MatProductSetFromOptions_MPIAIJ_ABt(C)); 2366 break; 2367 case MATPRODUCT_AtB: 2368 PetscCall(MatProductSetFromOptions_MPIAIJ_AtB(C)); 2369 break; 2370 case MATPRODUCT_PtAP: 2371 PetscCall(MatProductSetFromOptions_MPIAIJ_PtAP(C)); 2372 break; 2373 case MATPRODUCT_RARt: 2374 PetscCall(MatProductSetFromOptions_MPIAIJ_RARt(C)); 2375 break; 2376 case MATPRODUCT_ABC: 2377 PetscCall(MatProductSetFromOptions_MPIAIJ_ABC(C)); 2378 break; 2379 default: 2380 break; 2381 } 2382 PetscFunctionReturn(PETSC_SUCCESS); 2383 } 2384