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