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, *coa; 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 *dummy1, *dummy2, *dummy3, *dummy4; 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, &dummy1)); 143 PetscCall(MatSeqAIJGetArrayRead(a->B, &dummy2)); 144 PetscCall(MatSeqAIJGetArrayRead(ptap->P_loc, &dummy3)); 145 if (ptap->P_oth) PetscCall(MatSeqAIJGetArrayRead(ptap->P_oth, &dummy4)); 146 PetscCall(MatSeqAIJGetArrayWrite(c->A, &cda)); 147 PetscCall(MatSeqAIJGetArrayWrite(c->B, &coa)); 148 for (i = 0; i < cm; i++) { 149 /* compute apa = A[i,:]*P */ 150 AProw_nonscalable(i, ad, ao, p_loc, p_oth, apa); 151 152 /* set values in C */ 153 apJ = PetscSafePointerPlusOffset(apj, api[i]); 154 cdnz = cd->i[i + 1] - cd->i[i]; 155 conz = co->i[i + 1] - co->i[i]; 156 157 /* 1st off-diagonal part of C */ 158 ca = PetscSafePointerPlusOffset(coa, co->i[i]); 159 k = 0; 160 for (k0 = 0; k0 < conz; k0++) { 161 if (apJ[k] >= cstart) break; 162 ca[k0] = apa[apJ[k]]; 163 apa[apJ[k++]] = 0.0; 164 } 165 166 /* diagonal part of C */ 167 ca = PetscSafePointerPlusOffset(cda, cd->i[i]); 168 for (k1 = 0; k1 < cdnz; k1++) { 169 ca[k1] = apa[apJ[k]]; 170 apa[apJ[k++]] = 0.0; 171 } 172 173 /* 2nd off-diagonal part of C */ 174 ca = PetscSafePointerPlusOffset(coa, co->i[i]); 175 for (; k0 < conz; k0++) { 176 ca[k0] = apa[apJ[k]]; 177 apa[apJ[k++]] = 0.0; 178 } 179 } 180 PetscCall(MatSeqAIJRestoreArrayRead(a->A, &dummy1)); 181 PetscCall(MatSeqAIJRestoreArrayRead(a->B, &dummy2)); 182 PetscCall(MatSeqAIJRestoreArrayRead(ptap->P_loc, &dummy3)); 183 if (ptap->P_oth) PetscCall(MatSeqAIJRestoreArrayRead(ptap->P_oth, &dummy4)); 184 PetscCall(MatSeqAIJRestoreArrayWrite(c->A, &cda)); 185 PetscCall(MatSeqAIJRestoreArrayWrite(c->B, &coa)); 186 187 PetscCall(MatAssemblyBegin(C, MAT_FINAL_ASSEMBLY)); 188 PetscCall(MatAssemblyEnd(C, MAT_FINAL_ASSEMBLY)); 189 PetscFunctionReturn(PETSC_SUCCESS); 190 } 191 192 PetscErrorCode MatMatMultSymbolic_MPIAIJ_MPIAIJ_nonscalable(Mat A, Mat P, PetscReal fill, Mat C) 193 { 194 MPI_Comm comm; 195 PetscMPIInt size; 196 Mat_APMPI *ptap; 197 PetscFreeSpaceList free_space = NULL, current_space = NULL; 198 Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data; 199 Mat_SeqAIJ *ad = (Mat_SeqAIJ *)a->A->data, *ao = (Mat_SeqAIJ *)a->B->data, *p_loc, *p_oth; 200 PetscInt *pi_loc, *pj_loc, *pi_oth, *pj_oth, *dnz, *onz; 201 PetscInt *adi = ad->i, *adj = ad->j, *aoi = ao->i, *aoj = ao->j, rstart = A->rmap->rstart; 202 PetscInt *lnk, i, pnz, row, *api, *apj, *Jptr, apnz, nspacedouble = 0, j, nzi; 203 PetscInt am = A->rmap->n, pN = P->cmap->N, pn = P->cmap->n, pm = P->rmap->n; 204 PetscBT lnkbt; 205 PetscReal afill; 206 MatType mtype; 207 208 PetscFunctionBegin; 209 MatCheckProduct(C, 4); 210 PetscCheck(!C->product->data, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Extra product struct not empty"); 211 PetscCall(PetscObjectGetComm((PetscObject)A, &comm)); 212 PetscCallMPI(MPI_Comm_size(comm, &size)); 213 214 /* create struct Mat_APMPI and attached it to C later */ 215 PetscCall(PetscNew(&ptap)); 216 217 /* get P_oth by taking rows of P (= non-zero cols of local A) from other processors */ 218 PetscCall(MatGetBrowsOfAoCols_MPIAIJ(A, P, MAT_INITIAL_MATRIX, &ptap->startsj_s, &ptap->startsj_r, &ptap->bufa, &ptap->P_oth)); 219 220 /* get P_loc by taking all local rows of P */ 221 PetscCall(MatMPIAIJGetLocalMat(P, MAT_INITIAL_MATRIX, &ptap->P_loc)); 222 223 p_loc = (Mat_SeqAIJ *)ptap->P_loc->data; 224 pi_loc = p_loc->i; 225 pj_loc = p_loc->j; 226 if (size > 1) { 227 p_oth = (Mat_SeqAIJ *)ptap->P_oth->data; 228 pi_oth = p_oth->i; 229 pj_oth = p_oth->j; 230 } else { 231 p_oth = NULL; 232 pi_oth = NULL; 233 pj_oth = NULL; 234 } 235 236 /* first, compute symbolic AP = A_loc*P = A_diag*P_loc + A_off*P_oth */ 237 PetscCall(PetscMalloc1(am + 1, &api)); 238 ptap->api = api; 239 api[0] = 0; 240 241 /* create and initialize a linked list */ 242 PetscCall(PetscLLCondensedCreate(pN, pN, &lnk, &lnkbt)); 243 244 /* Initial FreeSpace size is fill*(nnz(A)+nnz(P)) */ 245 PetscCall(PetscFreeSpaceGet(PetscRealIntMultTruncate(fill, PetscIntSumTruncate(adi[am], PetscIntSumTruncate(aoi[am], pi_loc[pm]))), &free_space)); 246 current_space = free_space; 247 248 MatPreallocateBegin(comm, am, pn, dnz, onz); 249 for (i = 0; i < am; i++) { 250 /* diagonal portion of A */ 251 nzi = adi[i + 1] - adi[i]; 252 for (j = 0; j < nzi; j++) { 253 row = *adj++; 254 pnz = pi_loc[row + 1] - pi_loc[row]; 255 Jptr = pj_loc + pi_loc[row]; 256 /* add non-zero cols of P into the sorted linked list lnk */ 257 PetscCall(PetscLLCondensedAddSorted(pnz, Jptr, lnk, lnkbt)); 258 } 259 /* off-diagonal portion of A */ 260 nzi = aoi[i + 1] - aoi[i]; 261 for (j = 0; j < nzi; j++) { 262 row = *aoj++; 263 pnz = pi_oth[row + 1] - pi_oth[row]; 264 Jptr = pj_oth + pi_oth[row]; 265 PetscCall(PetscLLCondensedAddSorted(pnz, Jptr, lnk, lnkbt)); 266 } 267 /* add possible missing diagonal entry */ 268 if (C->force_diagonals) { 269 j = i + rstart; /* column index */ 270 PetscCall(PetscLLCondensedAddSorted(1, &j, lnk, lnkbt)); 271 } 272 273 apnz = lnk[0]; 274 api[i + 1] = api[i] + apnz; 275 276 /* if free space is not available, double the total space in the list */ 277 if (current_space->local_remaining < apnz) { 278 PetscCall(PetscFreeSpaceGet(PetscIntSumTruncate(apnz, current_space->total_array_size), ¤t_space)); 279 nspacedouble++; 280 } 281 282 /* Copy data into free space, then initialize lnk */ 283 PetscCall(PetscLLCondensedClean(pN, apnz, current_space->array, lnk, lnkbt)); 284 PetscCall(MatPreallocateSet(i + rstart, apnz, current_space->array, dnz, onz)); 285 286 current_space->array += apnz; 287 current_space->local_used += apnz; 288 current_space->local_remaining -= apnz; 289 } 290 291 /* Allocate space for apj, initialize apj, and */ 292 /* destroy list of free space and other temporary array(s) */ 293 PetscCall(PetscMalloc1(api[am], &ptap->apj)); 294 apj = ptap->apj; 295 PetscCall(PetscFreeSpaceContiguous(&free_space, ptap->apj)); 296 PetscCall(PetscLLDestroy(lnk, lnkbt)); 297 298 /* malloc apa to store dense row A[i,:]*P */ 299 PetscCall(PetscCalloc1(pN, &ptap->apa)); 300 301 /* set and assemble symbolic parallel matrix C */ 302 PetscCall(MatSetSizes(C, am, pn, PETSC_DETERMINE, PETSC_DETERMINE)); 303 PetscCall(MatSetBlockSizesFromMats(C, A, P)); 304 305 PetscCall(MatGetType(A, &mtype)); 306 PetscCall(MatSetType(C, mtype)); 307 PetscCall(MatMPIAIJSetPreallocation(C, 0, dnz, 0, onz)); 308 MatPreallocateEnd(dnz, onz); 309 310 PetscCall(MatSetValues_MPIAIJ_CopyFromCSRFormat_Symbolic(C, apj, api)); 311 PetscCall(MatAssemblyBegin(C, MAT_FINAL_ASSEMBLY)); 312 PetscCall(MatAssemblyEnd(C, MAT_FINAL_ASSEMBLY)); 313 PetscCall(MatSetOption(C, MAT_NEW_NONZERO_LOCATION_ERR, PETSC_TRUE)); 314 315 C->ops->matmultnumeric = MatMatMultNumeric_MPIAIJ_MPIAIJ_nonscalable; 316 C->ops->productnumeric = MatProductNumeric_AB; 317 318 /* attach the supporting struct to C for reuse */ 319 C->product->data = ptap; 320 C->product->destroy = MatDestroy_MPIAIJ_MatMatMult; 321 322 /* set MatInfo */ 323 afill = (PetscReal)api[am] / (adi[am] + aoi[am] + pi_loc[pm] + 1) + 1.e-5; 324 if (afill < 1.0) afill = 1.0; 325 C->info.mallocs = nspacedouble; 326 C->info.fill_ratio_given = fill; 327 C->info.fill_ratio_needed = afill; 328 329 #if defined(PETSC_USE_INFO) 330 if (api[am]) { 331 PetscCall(PetscInfo(C, "Reallocs %" PetscInt_FMT "; Fill ratio: given %g needed %g.\n", nspacedouble, (double)fill, (double)afill)); 332 PetscCall(PetscInfo(C, "Use MatMatMult(A,B,MatReuse,%g,&C) for best performance.;\n", (double)afill)); 333 } else { 334 PetscCall(PetscInfo(C, "Empty matrix product\n")); 335 } 336 #endif 337 PetscFunctionReturn(PETSC_SUCCESS); 338 } 339 340 static PetscErrorCode MatMatMultSymbolic_MPIAIJ_MPIDense(Mat, Mat, PetscReal, Mat); 341 static PetscErrorCode MatMatMultNumeric_MPIAIJ_MPIDense(Mat, Mat, Mat); 342 343 static PetscErrorCode MatProductSetFromOptions_MPIAIJ_MPIDense_AB(Mat C) 344 { 345 Mat_Product *product = C->product; 346 Mat A = product->A, B = product->B; 347 348 PetscFunctionBegin; 349 if (A->cmap->rstart != B->rmap->rstart || A->cmap->rend != B->rmap->rend) 350 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); 351 352 C->ops->matmultsymbolic = MatMatMultSymbolic_MPIAIJ_MPIDense; 353 C->ops->productsymbolic = MatProductSymbolic_AB; 354 PetscFunctionReturn(PETSC_SUCCESS); 355 } 356 357 static PetscErrorCode MatProductSetFromOptions_MPIAIJ_MPIDense_AtB(Mat C) 358 { 359 Mat_Product *product = C->product; 360 Mat A = product->A, B = product->B; 361 362 PetscFunctionBegin; 363 if (A->rmap->rstart != B->rmap->rstart || A->rmap->rend != B->rmap->rend) 364 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); 365 366 C->ops->transposematmultsymbolic = MatTransposeMatMultSymbolic_MPIAIJ_MPIDense; 367 C->ops->productsymbolic = MatProductSymbolic_AtB; 368 PetscFunctionReturn(PETSC_SUCCESS); 369 } 370 371 PETSC_INTERN PetscErrorCode MatProductSetFromOptions_MPIAIJ_MPIDense(Mat C) 372 { 373 Mat_Product *product = C->product; 374 375 PetscFunctionBegin; 376 switch (product->type) { 377 case MATPRODUCT_AB: 378 PetscCall(MatProductSetFromOptions_MPIAIJ_MPIDense_AB(C)); 379 break; 380 case MATPRODUCT_AtB: 381 PetscCall(MatProductSetFromOptions_MPIAIJ_MPIDense_AtB(C)); 382 break; 383 default: 384 break; 385 } 386 PetscFunctionReturn(PETSC_SUCCESS); 387 } 388 389 typedef struct { 390 Mat workB, workB1; 391 MPI_Request *rwaits, *swaits; 392 PetscInt nsends, nrecvs; 393 MPI_Datatype *stype, *rtype; 394 PetscInt blda; 395 } MPIAIJ_MPIDense; 396 397 static PetscErrorCode MatMPIAIJ_MPIDenseDestroy(void *ctx) 398 { 399 MPIAIJ_MPIDense *contents = (MPIAIJ_MPIDense *)ctx; 400 PetscInt i; 401 402 PetscFunctionBegin; 403 PetscCall(MatDestroy(&contents->workB)); 404 PetscCall(MatDestroy(&contents->workB1)); 405 for (i = 0; i < contents->nsends; i++) PetscCallMPI(MPI_Type_free(&contents->stype[i])); 406 for (i = 0; i < contents->nrecvs; i++) PetscCallMPI(MPI_Type_free(&contents->rtype[i])); 407 PetscCall(PetscFree4(contents->stype, contents->rtype, contents->rwaits, contents->swaits)); 408 PetscCall(PetscFree(contents)); 409 PetscFunctionReturn(PETSC_SUCCESS); 410 } 411 412 static PetscErrorCode MatMatMultSymbolic_MPIAIJ_MPIDense(Mat A, Mat B, PetscReal fill, Mat C) 413 { 414 Mat_MPIAIJ *aij = (Mat_MPIAIJ *)A->data; 415 PetscInt nz = aij->B->cmap->n, blda, m, M, n, N; 416 MPIAIJ_MPIDense *contents; 417 VecScatter ctx = aij->Mvctx; 418 PetscInt Am = A->rmap->n, Bm = B->rmap->n, BN = B->cmap->N, Bbn, Bbn1, bs, numBb; 419 MPI_Comm comm; 420 MPI_Datatype type1, *stype, *rtype; 421 const PetscInt *sindices, *sstarts, *rstarts; 422 PetscMPIInt *disp, nsends, nrecvs, nrows_to, nrows_from; 423 PetscBool cisdense; 424 425 PetscFunctionBegin; 426 MatCheckProduct(C, 4); 427 PetscCheck(!C->product->data, PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Product data not empty"); 428 PetscCall(PetscObjectGetComm((PetscObject)A, &comm)); 429 PetscCall(PetscObjectBaseTypeCompare((PetscObject)C, MATMPIDENSE, &cisdense)); 430 if (!cisdense) PetscCall(MatSetType(C, ((PetscObject)B)->type_name)); 431 PetscCall(MatGetLocalSize(C, &m, &n)); 432 PetscCall(MatGetSize(C, &M, &N)); 433 if (m == PETSC_DECIDE || n == PETSC_DECIDE || M == PETSC_DECIDE || N == PETSC_DECIDE) PetscCall(MatSetSizes(C, Am, B->cmap->n, A->rmap->N, BN)); 434 PetscCall(MatSetBlockSizesFromMats(C, A, B)); 435 PetscCall(MatSetUp(C)); 436 PetscCall(MatDenseGetLDA(B, &blda)); 437 PetscCall(PetscNew(&contents)); 438 439 PetscCall(VecScatterGetRemote_Private(ctx, PETSC_TRUE /*send*/, &nsends, &sstarts, &sindices, NULL, NULL)); 440 PetscCall(VecScatterGetRemoteOrdered_Private(ctx, PETSC_FALSE /*recv*/, &nrecvs, &rstarts, NULL, NULL, NULL)); 441 442 /* Create column block of B and C for memory scalability when BN is too large */ 443 /* Estimate Bbn, column size of Bb */ 444 if (nz) { 445 Bbn1 = 2 * Am * BN / nz; 446 if (!Bbn1) Bbn1 = 1; 447 } else Bbn1 = BN; 448 449 bs = B->cmap->bs; 450 Bbn1 = Bbn1 / bs * bs; /* Bbn1 is a multiple of bs */ 451 if (Bbn1 > BN) Bbn1 = BN; 452 PetscCallMPI(MPIU_Allreduce(&Bbn1, &Bbn, 1, MPIU_INT, MPI_MAX, comm)); 453 454 /* Enable runtime option for Bbn */ 455 PetscOptionsBegin(comm, ((PetscObject)C)->prefix, "MatMatMult", "Mat"); 456 PetscCall(PetscOptionsInt("-matmatmult_Bbn", "Number of columns in Bb", "MatMatMult", Bbn, &Bbn, NULL)); 457 PetscOptionsEnd(); 458 Bbn = PetscMin(Bbn, BN); 459 460 if (Bbn > 0 && Bbn < BN) { 461 numBb = BN / Bbn; 462 Bbn1 = BN - numBb * Bbn; 463 } else numBb = 0; 464 465 if (numBb) { 466 PetscCall(PetscInfo(C, "use Bb, BN=%" PetscInt_FMT ", Bbn=%" PetscInt_FMT "; numBb=%" PetscInt_FMT "\n", BN, Bbn, numBb)); 467 if (Bbn1) { /* Create workB1 for the remaining columns */ 468 PetscCall(PetscInfo(C, "use Bb1, BN=%" PetscInt_FMT ", Bbn1=%" PetscInt_FMT "\n", BN, Bbn1)); 469 /* Create work matrix used to store off processor rows of B needed for local product */ 470 PetscCall(MatCreateSeqDense(PETSC_COMM_SELF, nz, Bbn1, NULL, &contents->workB1)); 471 } else contents->workB1 = NULL; 472 } 473 474 /* Create work matrix used to store off processor rows of B needed for local product */ 475 PetscCall(MatCreateSeqDense(PETSC_COMM_SELF, nz, Bbn, NULL, &contents->workB)); 476 477 /* Use MPI derived data type to reduce memory required by the send/recv buffers */ 478 PetscCall(PetscMalloc4(nsends, &stype, nrecvs, &rtype, nrecvs, &contents->rwaits, nsends, &contents->swaits)); 479 contents->stype = stype; 480 contents->nsends = nsends; 481 482 contents->rtype = rtype; 483 contents->nrecvs = nrecvs; 484 contents->blda = blda; 485 486 PetscCall(PetscMalloc1(Bm + 1, &disp)); 487 for (PetscMPIInt i = 0; i < nsends; i++) { 488 PetscCall(PetscMPIIntCast(sstarts[i + 1] - sstarts[i], &nrows_to)); 489 for (PetscInt j = 0; j < nrows_to; j++) PetscCall(PetscMPIIntCast(sindices[sstarts[i] + j], &disp[j])); /* rowB to be sent */ 490 PetscCallMPI(MPI_Type_create_indexed_block(nrows_to, 1, disp, MPIU_SCALAR, &type1)); 491 PetscCallMPI(MPI_Type_create_resized(type1, 0, blda * sizeof(PetscScalar), &stype[i])); 492 PetscCallMPI(MPI_Type_commit(&stype[i])); 493 PetscCallMPI(MPI_Type_free(&type1)); 494 } 495 496 for (PetscMPIInt i = 0; i < nrecvs; i++) { 497 /* received values from a process form a (nrows_from x Bbn) row block in workB (column-wise) */ 498 PetscCall(PetscMPIIntCast(rstarts[i + 1] - rstarts[i], &nrows_from)); 499 disp[0] = 0; 500 PetscCallMPI(MPI_Type_create_indexed_block(1, nrows_from, disp, MPIU_SCALAR, &type1)); 501 PetscCallMPI(MPI_Type_create_resized(type1, 0, nz * sizeof(PetscScalar), &rtype[i])); 502 PetscCallMPI(MPI_Type_commit(&rtype[i])); 503 PetscCallMPI(MPI_Type_free(&type1)); 504 } 505 506 PetscCall(PetscFree(disp)); 507 PetscCall(VecScatterRestoreRemote_Private(ctx, PETSC_TRUE /*send*/, &nsends, &sstarts, &sindices, NULL, NULL)); 508 PetscCall(VecScatterRestoreRemoteOrdered_Private(ctx, PETSC_FALSE /*recv*/, &nrecvs, &rstarts, NULL, NULL, NULL)); 509 PetscCall(MatSetOption(C, MAT_NO_OFF_PROC_ENTRIES, PETSC_TRUE)); 510 PetscCall(MatAssemblyBegin(C, MAT_FINAL_ASSEMBLY)); 511 PetscCall(MatAssemblyEnd(C, MAT_FINAL_ASSEMBLY)); 512 513 C->product->data = contents; 514 C->product->destroy = MatMPIAIJ_MPIDenseDestroy; 515 C->ops->matmultnumeric = MatMatMultNumeric_MPIAIJ_MPIDense; 516 PetscFunctionReturn(PETSC_SUCCESS); 517 } 518 519 PETSC_INTERN PetscErrorCode MatMatMultNumericAdd_SeqAIJ_SeqDense(Mat, Mat, Mat, const PetscBool); 520 521 /* 522 Performs an efficient scatter on the rows of B needed by this process; this is 523 a modification of the VecScatterBegin_() routines. 524 525 Input: If Bbidx = 0, uses B = Bb, else B = Bb1, see MatMatMultSymbolic_MPIAIJ_MPIDense() 526 */ 527 528 static PetscErrorCode MatMPIDenseScatter(Mat A, Mat B, PetscInt Bbidx, Mat C, Mat *outworkB) 529 { 530 Mat_MPIAIJ *aij = (Mat_MPIAIJ *)A->data; 531 const PetscScalar *b; 532 PetscScalar *rvalues; 533 VecScatter ctx = aij->Mvctx; 534 const PetscInt *sindices, *sstarts, *rstarts; 535 const PetscMPIInt *sprocs, *rprocs; 536 PetscMPIInt nsends, nrecvs; 537 MPI_Request *swaits, *rwaits; 538 MPI_Comm comm; 539 PetscMPIInt tag = ((PetscObject)ctx)->tag, ncols, nrows, nsends_mpi, nrecvs_mpi; 540 MPIAIJ_MPIDense *contents; 541 Mat workB; 542 MPI_Datatype *stype, *rtype; 543 PetscInt blda; 544 545 PetscFunctionBegin; 546 MatCheckProduct(C, 4); 547 PetscCheck(C->product->data, PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Product data empty"); 548 PetscCall(PetscMPIIntCast(B->cmap->N, &ncols)); 549 PetscCall(PetscMPIIntCast(aij->B->cmap->n, &nrows)); 550 contents = (MPIAIJ_MPIDense *)C->product->data; 551 PetscCall(VecScatterGetRemote_Private(ctx, PETSC_TRUE /*send*/, &nsends, &sstarts, &sindices, &sprocs, NULL /*bs*/)); 552 PetscCall(VecScatterGetRemoteOrdered_Private(ctx, PETSC_FALSE /*recv*/, &nrecvs, &rstarts, NULL, &rprocs, NULL /*bs*/)); 553 PetscCall(PetscMPIIntCast(nsends, &nsends_mpi)); 554 PetscCall(PetscMPIIntCast(nrecvs, &nrecvs_mpi)); 555 if (Bbidx == 0) workB = *outworkB = contents->workB; 556 else workB = *outworkB = contents->workB1; 557 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); 558 swaits = contents->swaits; 559 rwaits = contents->rwaits; 560 561 PetscCall(MatDenseGetArrayRead(B, &b)); 562 PetscCall(MatDenseGetLDA(B, &blda)); 563 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); 564 PetscCall(MatDenseGetArray(workB, &rvalues)); 565 566 /* Post recv, use MPI derived data type to save memory */ 567 PetscCall(PetscObjectGetComm((PetscObject)C, &comm)); 568 rtype = contents->rtype; 569 for (PetscMPIInt i = 0; i < nrecvs; i++) PetscCallMPI(MPIU_Irecv(rvalues + (rstarts[i] - rstarts[0]), ncols, rtype[i], rprocs[i], tag, comm, rwaits + i)); 570 571 stype = contents->stype; 572 for (PetscMPIInt i = 0; i < nsends; i++) PetscCallMPI(MPIU_Isend(b, ncols, stype[i], sprocs[i], tag, comm, swaits + i)); 573 574 if (nrecvs) PetscCallMPI(MPI_Waitall(nrecvs_mpi, rwaits, MPI_STATUSES_IGNORE)); 575 if (nsends) PetscCallMPI(MPI_Waitall(nsends_mpi, swaits, MPI_STATUSES_IGNORE)); 576 577 PetscCall(VecScatterRestoreRemote_Private(ctx, PETSC_TRUE /*send*/, &nsends, &sstarts, &sindices, &sprocs, NULL)); 578 PetscCall(VecScatterRestoreRemoteOrdered_Private(ctx, PETSC_FALSE /*recv*/, &nrecvs, &rstarts, NULL, &rprocs, NULL)); 579 PetscCall(MatDenseRestoreArrayRead(B, &b)); 580 PetscCall(MatDenseRestoreArray(workB, &rvalues)); 581 PetscFunctionReturn(PETSC_SUCCESS); 582 } 583 584 static PetscErrorCode MatMatMultNumeric_MPIAIJ_MPIDense(Mat A, Mat B, Mat C) 585 { 586 Mat_MPIAIJ *aij = (Mat_MPIAIJ *)A->data; 587 Mat_MPIDense *bdense = (Mat_MPIDense *)B->data; 588 Mat_MPIDense *cdense = (Mat_MPIDense *)C->data; 589 Mat workB; 590 MPIAIJ_MPIDense *contents; 591 592 PetscFunctionBegin; 593 MatCheckProduct(C, 3); 594 PetscCheck(C->product->data, PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Product data empty"); 595 contents = (MPIAIJ_MPIDense *)C->product->data; 596 /* diagonal block of A times all local rows of B */ 597 /* TODO: this calls a symbolic multiplication every time, which could be avoided */ 598 PetscCall(MatMatMult(aij->A, bdense->A, MAT_REUSE_MATRIX, PETSC_CURRENT, &cdense->A)); 599 if (contents->workB->cmap->n == B->cmap->N) { 600 /* get off processor parts of B needed to complete C=A*B */ 601 PetscCall(MatMPIDenseScatter(A, B, 0, C, &workB)); 602 603 /* off-diagonal block of A times nonlocal rows of B */ 604 PetscCall(MatMatMultNumericAdd_SeqAIJ_SeqDense(aij->B, workB, cdense->A, PETSC_TRUE)); 605 } else { 606 Mat Bb, Cb; 607 PetscInt BN = B->cmap->N, n = contents->workB->cmap->n; 608 PetscBool ccpu; 609 610 PetscCheck(n > 0, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Column block size %" PetscInt_FMT " must be positive", n); 611 /* Prevent from unneeded copies back and forth from the GPU 612 when getting and restoring the submatrix 613 We need a proper GPU code for AIJ * dense in parallel */ 614 PetscCall(MatBoundToCPU(C, &ccpu)); 615 PetscCall(MatBindToCPU(C, PETSC_TRUE)); 616 for (PetscInt i = 0; i < BN; i += n) { 617 PetscCall(MatDenseGetSubMatrix(B, PETSC_DECIDE, PETSC_DECIDE, i, PetscMin(i + n, BN), &Bb)); 618 PetscCall(MatDenseGetSubMatrix(C, PETSC_DECIDE, PETSC_DECIDE, i, PetscMin(i + n, BN), &Cb)); 619 620 /* get off processor parts of B needed to complete C=A*B */ 621 PetscCall(MatMPIDenseScatter(A, Bb, (i + n) > BN, C, &workB)); 622 623 /* off-diagonal block of A times nonlocal rows of B */ 624 cdense = (Mat_MPIDense *)Cb->data; 625 PetscCall(MatMatMultNumericAdd_SeqAIJ_SeqDense(aij->B, workB, cdense->A, PETSC_TRUE)); 626 PetscCall(MatDenseRestoreSubMatrix(B, &Bb)); 627 PetscCall(MatDenseRestoreSubMatrix(C, &Cb)); 628 } 629 PetscCall(MatBindToCPU(C, ccpu)); 630 } 631 PetscFunctionReturn(PETSC_SUCCESS); 632 } 633 634 PetscErrorCode MatMatMultNumeric_MPIAIJ_MPIAIJ(Mat A, Mat P, Mat C) 635 { 636 Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data, *c = (Mat_MPIAIJ *)C->data; 637 Mat_SeqAIJ *ad = (Mat_SeqAIJ *)a->A->data, *ao = (Mat_SeqAIJ *)a->B->data; 638 Mat_SeqAIJ *cd = (Mat_SeqAIJ *)c->A->data, *co = (Mat_SeqAIJ *)c->B->data; 639 PetscInt *adi = ad->i, *adj, *aoi = ao->i, *aoj; 640 PetscScalar *ada, *aoa, *cda = cd->a, *coa = co->a; 641 Mat_SeqAIJ *p_loc, *p_oth; 642 PetscInt *pi_loc, *pj_loc, *pi_oth, *pj_oth, *pj; 643 PetscScalar *pa_loc, *pa_oth, *pa, valtmp, *ca; 644 PetscInt cm = C->rmap->n, anz, pnz; 645 Mat_APMPI *ptap; 646 PetscScalar *apa_sparse; 647 const PetscScalar *dummy; 648 PetscInt *api, *apj, *apJ, i, j, k, row; 649 PetscInt cstart = C->cmap->rstart; 650 PetscInt cdnz, conz, k0, k1, nextp; 651 MPI_Comm comm; 652 PetscMPIInt size; 653 654 PetscFunctionBegin; 655 MatCheckProduct(C, 3); 656 ptap = (Mat_APMPI *)C->product->data; 657 PetscCheck(ptap, PetscObjectComm((PetscObject)C), PETSC_ERR_ARG_WRONGSTATE, "PtAP cannot be computed. Missing data"); 658 PetscCall(PetscObjectGetComm((PetscObject)C, &comm)); 659 PetscCallMPI(MPI_Comm_size(comm, &size)); 660 PetscCheck(ptap->P_oth || size <= 1, PetscObjectComm((PetscObject)C), PETSC_ERR_ARG_WRONGSTATE, "AP cannot be reused. Do not call MatProductClear()"); 661 662 /* flag CPU mask for C */ 663 #if defined(PETSC_HAVE_DEVICE) 664 if (C->offloadmask != PETSC_OFFLOAD_UNALLOCATED) C->offloadmask = PETSC_OFFLOAD_CPU; 665 if (c->A->offloadmask != PETSC_OFFLOAD_UNALLOCATED) c->A->offloadmask = PETSC_OFFLOAD_CPU; 666 if (c->B->offloadmask != PETSC_OFFLOAD_UNALLOCATED) c->B->offloadmask = PETSC_OFFLOAD_CPU; 667 #endif 668 apa_sparse = ptap->apa; 669 670 /* 1) get P_oth = ptap->P_oth and P_loc = ptap->P_loc */ 671 /* update numerical values of P_oth and P_loc */ 672 PetscCall(MatGetBrowsOfAoCols_MPIAIJ(A, P, MAT_REUSE_MATRIX, &ptap->startsj_s, &ptap->startsj_r, &ptap->bufa, &ptap->P_oth)); 673 PetscCall(MatMPIAIJGetLocalMat(P, MAT_REUSE_MATRIX, &ptap->P_loc)); 674 675 /* 2) compute numeric C_loc = A_loc*P = Ad*P_loc + Ao*P_oth */ 676 /* get data from symbolic products */ 677 p_loc = (Mat_SeqAIJ *)ptap->P_loc->data; 678 pi_loc = p_loc->i; 679 pj_loc = p_loc->j; 680 pa_loc = p_loc->a; 681 if (size > 1) { 682 p_oth = (Mat_SeqAIJ *)ptap->P_oth->data; 683 pi_oth = p_oth->i; 684 pj_oth = p_oth->j; 685 pa_oth = p_oth->a; 686 } else { 687 p_oth = NULL; 688 pi_oth = NULL; 689 pj_oth = NULL; 690 pa_oth = NULL; 691 } 692 693 /* trigger copy to CPU */ 694 PetscCall(MatSeqAIJGetArrayRead(a->A, &dummy)); 695 PetscCall(MatSeqAIJRestoreArrayRead(a->A, &dummy)); 696 PetscCall(MatSeqAIJGetArrayRead(a->B, &dummy)); 697 PetscCall(MatSeqAIJRestoreArrayRead(a->B, &dummy)); 698 api = ptap->api; 699 apj = ptap->apj; 700 for (i = 0; i < cm; i++) { 701 apJ = apj + api[i]; 702 703 /* diagonal portion of A */ 704 anz = adi[i + 1] - adi[i]; 705 adj = ad->j + adi[i]; 706 ada = ad->a + adi[i]; 707 for (j = 0; j < anz; j++) { 708 row = adj[j]; 709 pnz = pi_loc[row + 1] - pi_loc[row]; 710 pj = pj_loc + pi_loc[row]; 711 pa = pa_loc + pi_loc[row]; 712 /* perform sparse axpy */ 713 valtmp = ada[j]; 714 nextp = 0; 715 for (k = 0; nextp < pnz; k++) { 716 if (apJ[k] == pj[nextp]) { /* column of AP == column of P */ 717 apa_sparse[k] += valtmp * pa[nextp++]; 718 } 719 } 720 PetscCall(PetscLogFlops(2.0 * pnz)); 721 } 722 723 /* off-diagonal portion of A */ 724 anz = aoi[i + 1] - aoi[i]; 725 aoj = PetscSafePointerPlusOffset(ao->j, aoi[i]); 726 aoa = PetscSafePointerPlusOffset(ao->a, aoi[i]); 727 for (j = 0; j < anz; j++) { 728 row = aoj[j]; 729 pnz = pi_oth[row + 1] - pi_oth[row]; 730 pj = pj_oth + pi_oth[row]; 731 pa = pa_oth + pi_oth[row]; 732 /* perform sparse axpy */ 733 valtmp = aoa[j]; 734 nextp = 0; 735 for (k = 0; nextp < pnz; k++) { 736 if (apJ[k] == pj[nextp]) { /* column of AP == column of P */ 737 apa_sparse[k] += valtmp * pa[nextp++]; 738 } 739 } 740 PetscCall(PetscLogFlops(2.0 * pnz)); 741 } 742 743 /* set values in C */ 744 cdnz = cd->i[i + 1] - cd->i[i]; 745 conz = co->i[i + 1] - co->i[i]; 746 747 /* 1st off-diagonal part of C */ 748 ca = PetscSafePointerPlusOffset(coa, co->i[i]); 749 k = 0; 750 for (k0 = 0; k0 < conz; k0++) { 751 if (apJ[k] >= cstart) break; 752 ca[k0] = apa_sparse[k]; 753 apa_sparse[k] = 0.0; 754 k++; 755 } 756 757 /* diagonal part of C */ 758 ca = cda + cd->i[i]; 759 for (k1 = 0; k1 < cdnz; k1++) { 760 ca[k1] = apa_sparse[k]; 761 apa_sparse[k] = 0.0; 762 k++; 763 } 764 765 /* 2nd off-diagonal part of C */ 766 ca = PetscSafePointerPlusOffset(coa, co->i[i]); 767 for (; k0 < conz; k0++) { 768 ca[k0] = apa_sparse[k]; 769 apa_sparse[k] = 0.0; 770 k++; 771 } 772 } 773 PetscCall(MatAssemblyBegin(C, MAT_FINAL_ASSEMBLY)); 774 PetscCall(MatAssemblyEnd(C, MAT_FINAL_ASSEMBLY)); 775 PetscFunctionReturn(PETSC_SUCCESS); 776 } 777 778 /* same as MatMatMultSymbolic_MPIAIJ_MPIAIJ_nonscalable(), except using LLCondensed to avoid O(BN) memory requirement */ 779 PetscErrorCode MatMatMultSymbolic_MPIAIJ_MPIAIJ(Mat A, Mat P, PetscReal fill, Mat C) 780 { 781 MPI_Comm comm; 782 PetscMPIInt size; 783 Mat_APMPI *ptap; 784 PetscFreeSpaceList free_space = NULL, current_space = NULL; 785 Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data; 786 Mat_SeqAIJ *ad = (Mat_SeqAIJ *)a->A->data, *ao = (Mat_SeqAIJ *)a->B->data, *p_loc, *p_oth; 787 PetscInt *pi_loc, *pj_loc, *pi_oth, *pj_oth, *dnz, *onz; 788 PetscInt *adi = ad->i, *adj = ad->j, *aoi = ao->i, *aoj = ao->j, rstart = A->rmap->rstart; 789 PetscInt i, pnz, row, *api, *apj, *Jptr, apnz, nspacedouble = 0, j, nzi, *lnk, apnz_max = 1; 790 PetscInt am = A->rmap->n, pn = P->cmap->n, pm = P->rmap->n, lsize = pn + 20; 791 PetscReal afill; 792 MatType mtype; 793 794 PetscFunctionBegin; 795 MatCheckProduct(C, 4); 796 PetscCheck(!C->product->data, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Extra product struct not empty"); 797 PetscCall(PetscObjectGetComm((PetscObject)A, &comm)); 798 PetscCallMPI(MPI_Comm_size(comm, &size)); 799 800 /* create struct Mat_APMPI and attached it to C later */ 801 PetscCall(PetscNew(&ptap)); 802 803 /* get P_oth by taking rows of P (= non-zero cols of local A) from other processors */ 804 PetscCall(MatGetBrowsOfAoCols_MPIAIJ(A, P, MAT_INITIAL_MATRIX, &ptap->startsj_s, &ptap->startsj_r, &ptap->bufa, &ptap->P_oth)); 805 806 /* get P_loc by taking all local rows of P */ 807 PetscCall(MatMPIAIJGetLocalMat(P, MAT_INITIAL_MATRIX, &ptap->P_loc)); 808 809 p_loc = (Mat_SeqAIJ *)ptap->P_loc->data; 810 pi_loc = p_loc->i; 811 pj_loc = p_loc->j; 812 if (size > 1) { 813 p_oth = (Mat_SeqAIJ *)ptap->P_oth->data; 814 pi_oth = p_oth->i; 815 pj_oth = p_oth->j; 816 } else { 817 p_oth = NULL; 818 pi_oth = NULL; 819 pj_oth = NULL; 820 } 821 822 /* first, compute symbolic AP = A_loc*P = A_diag*P_loc + A_off*P_oth */ 823 PetscCall(PetscMalloc1(am + 1, &api)); 824 ptap->api = api; 825 api[0] = 0; 826 827 PetscCall(PetscLLCondensedCreate_Scalable(lsize, &lnk)); 828 829 /* Initial FreeSpace size is fill*(nnz(A)+nnz(P)) */ 830 PetscCall(PetscFreeSpaceGet(PetscRealIntMultTruncate(fill, PetscIntSumTruncate(adi[am], PetscIntSumTruncate(aoi[am], pi_loc[pm]))), &free_space)); 831 current_space = free_space; 832 MatPreallocateBegin(comm, am, pn, dnz, onz); 833 for (i = 0; i < am; i++) { 834 /* diagonal portion of A */ 835 nzi = adi[i + 1] - adi[i]; 836 for (j = 0; j < nzi; j++) { 837 row = *adj++; 838 pnz = pi_loc[row + 1] - pi_loc[row]; 839 Jptr = pj_loc + pi_loc[row]; 840 /* Expand list if it is not long enough */ 841 if (pnz + apnz_max > lsize) { 842 lsize = pnz + apnz_max; 843 PetscCall(PetscLLCondensedExpand_Scalable(lsize, &lnk)); 844 } 845 /* add non-zero cols of P into the sorted linked list lnk */ 846 PetscCall(PetscLLCondensedAddSorted_Scalable(pnz, Jptr, lnk)); 847 apnz = *lnk; /* The first element in the list is the number of items in the list */ 848 api[i + 1] = api[i] + apnz; 849 if (apnz > apnz_max) apnz_max = apnz + 1; /* '1' for diagonal entry */ 850 } 851 /* off-diagonal portion of A */ 852 nzi = aoi[i + 1] - aoi[i]; 853 for (j = 0; j < nzi; j++) { 854 row = *aoj++; 855 pnz = pi_oth[row + 1] - pi_oth[row]; 856 Jptr = pj_oth + pi_oth[row]; 857 /* Expand list if it is not long enough */ 858 if (pnz + apnz_max > lsize) { 859 lsize = pnz + apnz_max; 860 PetscCall(PetscLLCondensedExpand_Scalable(lsize, &lnk)); 861 } 862 /* add non-zero cols of P into the sorted linked list lnk */ 863 PetscCall(PetscLLCondensedAddSorted_Scalable(pnz, Jptr, lnk)); 864 apnz = *lnk; /* The first element in the list is the number of items in the list */ 865 api[i + 1] = api[i] + apnz; 866 if (apnz > apnz_max) apnz_max = apnz + 1; /* '1' for diagonal entry */ 867 } 868 869 /* add missing diagonal entry */ 870 if (C->force_diagonals) { 871 j = i + rstart; /* column index */ 872 PetscCall(PetscLLCondensedAddSorted_Scalable(1, &j, lnk)); 873 } 874 875 apnz = *lnk; 876 api[i + 1] = api[i] + apnz; 877 if (apnz > apnz_max) apnz_max = apnz; 878 879 /* if free space is not available, double the total space in the list */ 880 if (current_space->local_remaining < apnz) { 881 PetscCall(PetscFreeSpaceGet(PetscIntSumTruncate(apnz, current_space->total_array_size), ¤t_space)); 882 nspacedouble++; 883 } 884 885 /* Copy data into free space, then initialize lnk */ 886 PetscCall(PetscLLCondensedClean_Scalable(apnz, current_space->array, lnk)); 887 PetscCall(MatPreallocateSet(i + rstart, apnz, current_space->array, dnz, onz)); 888 889 current_space->array += apnz; 890 current_space->local_used += apnz; 891 current_space->local_remaining -= apnz; 892 } 893 894 /* Allocate space for apj, initialize apj, and */ 895 /* destroy list of free space and other temporary array(s) */ 896 PetscCall(PetscMalloc1(api[am], &ptap->apj)); 897 apj = ptap->apj; 898 PetscCall(PetscFreeSpaceContiguous(&free_space, ptap->apj)); 899 PetscCall(PetscLLCondensedDestroy_Scalable(lnk)); 900 901 /* create and assemble symbolic parallel matrix C */ 902 PetscCall(MatSetSizes(C, am, pn, PETSC_DETERMINE, PETSC_DETERMINE)); 903 PetscCall(MatSetBlockSizesFromMats(C, A, P)); 904 PetscCall(MatGetType(A, &mtype)); 905 PetscCall(MatSetType(C, mtype)); 906 PetscCall(MatMPIAIJSetPreallocation(C, 0, dnz, 0, onz)); 907 MatPreallocateEnd(dnz, onz); 908 909 /* malloc apa for assembly C */ 910 PetscCall(PetscCalloc1(apnz_max, &ptap->apa)); 911 912 PetscCall(MatSetValues_MPIAIJ_CopyFromCSRFormat_Symbolic(C, apj, api)); 913 PetscCall(MatAssemblyBegin(C, MAT_FINAL_ASSEMBLY)); 914 PetscCall(MatAssemblyEnd(C, MAT_FINAL_ASSEMBLY)); 915 PetscCall(MatSetOption(C, MAT_NEW_NONZERO_LOCATION_ERR, PETSC_TRUE)); 916 917 C->ops->matmultnumeric = MatMatMultNumeric_MPIAIJ_MPIAIJ; 918 C->ops->productnumeric = MatProductNumeric_AB; 919 920 /* attach the supporting struct to C for reuse */ 921 C->product->data = ptap; 922 C->product->destroy = MatDestroy_MPIAIJ_MatMatMult; 923 924 /* set MatInfo */ 925 afill = (PetscReal)api[am] / (adi[am] + aoi[am] + pi_loc[pm] + 1) + 1.e-5; 926 if (afill < 1.0) afill = 1.0; 927 C->info.mallocs = nspacedouble; 928 C->info.fill_ratio_given = fill; 929 C->info.fill_ratio_needed = afill; 930 931 #if defined(PETSC_USE_INFO) 932 if (api[am]) { 933 PetscCall(PetscInfo(C, "Reallocs %" PetscInt_FMT "; Fill ratio: given %g needed %g.\n", nspacedouble, (double)fill, (double)afill)); 934 PetscCall(PetscInfo(C, "Use MatMatMult(A,B,MatReuse,%g,&C) for best performance.;\n", (double)afill)); 935 } else { 936 PetscCall(PetscInfo(C, "Empty matrix product\n")); 937 } 938 #endif 939 PetscFunctionReturn(PETSC_SUCCESS); 940 } 941 942 /* This function is needed for the seqMPI matrix-matrix multiplication. */ 943 /* Three input arrays are merged to one output array. The size of the */ 944 /* output array is also output. Duplicate entries only show up once. */ 945 static void Merge3SortedArrays(PetscInt size1, PetscInt *in1, PetscInt size2, PetscInt *in2, PetscInt size3, PetscInt *in3, PetscInt *size4, PetscInt *out) 946 { 947 int i = 0, j = 0, k = 0, l = 0; 948 949 /* Traverse all three arrays */ 950 while (i < size1 && j < size2 && k < size3) { 951 if (in1[i] < in2[j] && in1[i] < in3[k]) { 952 out[l++] = in1[i++]; 953 } else if (in2[j] < in1[i] && in2[j] < in3[k]) { 954 out[l++] = in2[j++]; 955 } else if (in3[k] < in1[i] && in3[k] < in2[j]) { 956 out[l++] = in3[k++]; 957 } else if (in1[i] == in2[j] && in1[i] < in3[k]) { 958 out[l++] = in1[i]; 959 i++, j++; 960 } else if (in1[i] == in3[k] && in1[i] < in2[j]) { 961 out[l++] = in1[i]; 962 i++, k++; 963 } else if (in3[k] == in2[j] && in2[j] < in1[i]) { 964 out[l++] = in2[j]; 965 k++, j++; 966 } else if (in1[i] == in2[j] && in1[i] == in3[k]) { 967 out[l++] = in1[i]; 968 i++, j++, k++; 969 } 970 } 971 972 /* Traverse two remaining arrays */ 973 while (i < size1 && j < size2) { 974 if (in1[i] < in2[j]) { 975 out[l++] = in1[i++]; 976 } else if (in1[i] > in2[j]) { 977 out[l++] = in2[j++]; 978 } else { 979 out[l++] = in1[i]; 980 i++, j++; 981 } 982 } 983 984 while (i < size1 && k < size3) { 985 if (in1[i] < in3[k]) { 986 out[l++] = in1[i++]; 987 } else if (in1[i] > in3[k]) { 988 out[l++] = in3[k++]; 989 } else { 990 out[l++] = in1[i]; 991 i++, k++; 992 } 993 } 994 995 while (k < size3 && j < size2) { 996 if (in3[k] < in2[j]) { 997 out[l++] = in3[k++]; 998 } else if (in3[k] > in2[j]) { 999 out[l++] = in2[j++]; 1000 } else { 1001 out[l++] = in3[k]; 1002 k++, j++; 1003 } 1004 } 1005 1006 /* Traverse one remaining array */ 1007 while (i < size1) out[l++] = in1[i++]; 1008 while (j < size2) out[l++] = in2[j++]; 1009 while (k < size3) out[l++] = in3[k++]; 1010 1011 *size4 = l; 1012 } 1013 1014 /* This matrix-matrix multiplication algorithm divides the multiplication into three multiplications and */ 1015 /* adds up the products. Two of these three multiplications are performed with existing (sequential) */ 1016 /* matrix-matrix multiplications. */ 1017 PetscErrorCode MatMatMultSymbolic_MPIAIJ_MPIAIJ_seqMPI(Mat A, Mat P, PetscReal fill, Mat C) 1018 { 1019 MPI_Comm comm; 1020 PetscMPIInt size; 1021 Mat_APMPI *ptap; 1022 PetscFreeSpaceList free_space_diag = NULL, current_space = NULL; 1023 Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data; 1024 Mat_SeqAIJ *ad = (Mat_SeqAIJ *)a->A->data, *ao = (Mat_SeqAIJ *)a->B->data, *p_loc; 1025 Mat_MPIAIJ *p = (Mat_MPIAIJ *)P->data; 1026 Mat_SeqAIJ *adpd_seq, *p_off, *aopoth_seq; 1027 PetscInt adponz, adpdnz; 1028 PetscInt *pi_loc, *dnz, *onz; 1029 PetscInt *adi = ad->i, *adj = ad->j, *aoi = ao->i, rstart = A->rmap->rstart; 1030 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; 1031 PetscInt am = A->rmap->n, pN = P->cmap->N, pn = P->cmap->n, pm = P->rmap->n, p_colstart, p_colend; 1032 PetscBT lnkbt; 1033 PetscReal afill; 1034 PetscMPIInt rank; 1035 Mat adpd, aopoth; 1036 MatType mtype; 1037 const char *prefix; 1038 1039 PetscFunctionBegin; 1040 MatCheckProduct(C, 4); 1041 PetscCheck(!C->product->data, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Extra product struct not empty"); 1042 PetscCall(PetscObjectGetComm((PetscObject)A, &comm)); 1043 PetscCallMPI(MPI_Comm_size(comm, &size)); 1044 PetscCallMPI(MPI_Comm_rank(comm, &rank)); 1045 PetscCall(MatGetOwnershipRangeColumn(P, &p_colstart, &p_colend)); 1046 1047 /* create struct Mat_APMPI and attached it to C later */ 1048 PetscCall(PetscNew(&ptap)); 1049 1050 /* get P_oth by taking rows of P (= non-zero cols of local A) from other processors */ 1051 PetscCall(MatGetBrowsOfAoCols_MPIAIJ(A, P, MAT_INITIAL_MATRIX, &ptap->startsj_s, &ptap->startsj_r, &ptap->bufa, &ptap->P_oth)); 1052 1053 /* get P_loc by taking all local rows of P */ 1054 PetscCall(MatMPIAIJGetLocalMat(P, MAT_INITIAL_MATRIX, &ptap->P_loc)); 1055 1056 p_loc = (Mat_SeqAIJ *)ptap->P_loc->data; 1057 pi_loc = p_loc->i; 1058 1059 /* Allocate memory for the i arrays of the matrices A*P, A_diag*P_off and A_offd * P */ 1060 PetscCall(PetscMalloc1(am + 1, &api)); 1061 PetscCall(PetscMalloc1(am + 1, &adpoi)); 1062 1063 adpoi[0] = 0; 1064 ptap->api = api; 1065 api[0] = 0; 1066 1067 /* create and initialize a linked list, will be used for both A_diag * P_loc_off and A_offd * P_oth */ 1068 PetscCall(PetscLLCondensedCreate(pN, pN, &lnk, &lnkbt)); 1069 MatPreallocateBegin(comm, am, pn, dnz, onz); 1070 1071 /* Symbolic calc of A_loc_diag * P_loc_diag */ 1072 PetscCall(MatGetOptionsPrefix(A, &prefix)); 1073 PetscCall(MatProductCreate(a->A, p->A, NULL, &adpd)); 1074 PetscCall(MatGetOptionsPrefix(A, &prefix)); 1075 PetscCall(MatSetOptionsPrefix(adpd, prefix)); 1076 PetscCall(MatAppendOptionsPrefix(adpd, "inner_diag_")); 1077 1078 PetscCall(MatProductSetType(adpd, MATPRODUCT_AB)); 1079 PetscCall(MatProductSetAlgorithm(adpd, "sorted")); 1080 PetscCall(MatProductSetFill(adpd, fill)); 1081 PetscCall(MatProductSetFromOptions(adpd)); 1082 1083 adpd->force_diagonals = C->force_diagonals; 1084 PetscCall(MatProductSymbolic(adpd)); 1085 1086 adpd_seq = (Mat_SeqAIJ *)((adpd)->data); 1087 adpdi = adpd_seq->i; 1088 adpdj = adpd_seq->j; 1089 p_off = (Mat_SeqAIJ *)p->B->data; 1090 poff_i = p_off->i; 1091 poff_j = p_off->j; 1092 1093 /* j_temp stores indices of a result row before they are added to the linked list */ 1094 PetscCall(PetscMalloc1(pN, &j_temp)); 1095 1096 /* Symbolic calc of the A_diag * p_loc_off */ 1097 /* Initial FreeSpace size is fill*(nnz(A)+nnz(P)) */ 1098 PetscCall(PetscFreeSpaceGet(PetscRealIntMultTruncate(fill, PetscIntSumTruncate(adi[am], PetscIntSumTruncate(aoi[am], pi_loc[pm]))), &free_space_diag)); 1099 current_space = free_space_diag; 1100 1101 for (i = 0; i < am; i++) { 1102 /* A_diag * P_loc_off */ 1103 nzi = adi[i + 1] - adi[i]; 1104 for (j = 0; j < nzi; j++) { 1105 row = *adj++; 1106 pnz = poff_i[row + 1] - poff_i[row]; 1107 Jptr = poff_j + poff_i[row]; 1108 for (i1 = 0; i1 < pnz; i1++) j_temp[i1] = p->garray[Jptr[i1]]; 1109 /* add non-zero cols of P into the sorted linked list lnk */ 1110 PetscCall(PetscLLCondensedAddSorted(pnz, j_temp, lnk, lnkbt)); 1111 } 1112 1113 adponz = lnk[0]; 1114 adpoi[i + 1] = adpoi[i] + adponz; 1115 1116 /* if free space is not available, double the total space in the list */ 1117 if (current_space->local_remaining < adponz) { 1118 PetscCall(PetscFreeSpaceGet(PetscIntSumTruncate(adponz, current_space->total_array_size), ¤t_space)); 1119 nspacedouble++; 1120 } 1121 1122 /* Copy data into free space, then initialize lnk */ 1123 PetscCall(PetscLLCondensedClean(pN, adponz, current_space->array, lnk, lnkbt)); 1124 1125 current_space->array += adponz; 1126 current_space->local_used += adponz; 1127 current_space->local_remaining -= adponz; 1128 } 1129 1130 /* Symbolic calc of A_off * P_oth */ 1131 PetscCall(MatSetOptionsPrefix(a->B, prefix)); 1132 PetscCall(MatAppendOptionsPrefix(a->B, "inner_offdiag_")); 1133 PetscCall(MatCreate(PETSC_COMM_SELF, &aopoth)); 1134 PetscCall(MatMatMultSymbolic_SeqAIJ_SeqAIJ(a->B, ptap->P_oth, fill, aopoth)); 1135 aopoth_seq = (Mat_SeqAIJ *)((aopoth)->data); 1136 aopothi = aopoth_seq->i; 1137 aopothj = aopoth_seq->j; 1138 1139 /* Allocate space for apj, adpj, aopj, ... */ 1140 /* destroy lists of free space and other temporary array(s) */ 1141 1142 PetscCall(PetscMalloc1(aopothi[am] + adpoi[am] + adpdi[am], &ptap->apj)); 1143 PetscCall(PetscMalloc1(adpoi[am], &adpoj)); 1144 1145 /* Copy from linked list to j-array */ 1146 PetscCall(PetscFreeSpaceContiguous(&free_space_diag, adpoj)); 1147 PetscCall(PetscLLDestroy(lnk, lnkbt)); 1148 1149 adpoJ = adpoj; 1150 adpdJ = adpdj; 1151 aopJ = aopothj; 1152 apj = ptap->apj; 1153 apJ = apj; /* still empty */ 1154 1155 /* Merge j-arrays of A_off * P, A_diag * P_loc_off, and */ 1156 /* A_diag * P_loc_diag to get A*P */ 1157 for (i = 0; i < am; i++) { 1158 aopnz = aopothi[i + 1] - aopothi[i]; 1159 adponz = adpoi[i + 1] - adpoi[i]; 1160 adpdnz = adpdi[i + 1] - adpdi[i]; 1161 1162 /* Correct indices from A_diag*P_diag */ 1163 for (i1 = 0; i1 < adpdnz; i1++) adpdJ[i1] += p_colstart; 1164 /* Merge j-arrays of A_diag * P_loc_off and A_diag * P_loc_diag and A_off * P_oth */ 1165 Merge3SortedArrays(adponz, adpoJ, adpdnz, adpdJ, aopnz, aopJ, &apnz, apJ); 1166 PetscCall(MatPreallocateSet(i + rstart, apnz, apJ, dnz, onz)); 1167 1168 aopJ += aopnz; 1169 adpoJ += adponz; 1170 adpdJ += adpdnz; 1171 apJ += apnz; 1172 api[i + 1] = api[i] + apnz; 1173 } 1174 1175 /* malloc apa to store dense row A[i,:]*P */ 1176 PetscCall(PetscCalloc1(pN, &ptap->apa)); 1177 1178 /* create and assemble symbolic parallel matrix C */ 1179 PetscCall(MatSetSizes(C, am, pn, PETSC_DETERMINE, PETSC_DETERMINE)); 1180 PetscCall(MatSetBlockSizesFromMats(C, A, P)); 1181 PetscCall(MatGetType(A, &mtype)); 1182 PetscCall(MatSetType(C, mtype)); 1183 PetscCall(MatMPIAIJSetPreallocation(C, 0, dnz, 0, onz)); 1184 MatPreallocateEnd(dnz, onz); 1185 1186 PetscCall(MatSetValues_MPIAIJ_CopyFromCSRFormat_Symbolic(C, apj, api)); 1187 PetscCall(MatAssemblyBegin(C, MAT_FINAL_ASSEMBLY)); 1188 PetscCall(MatAssemblyEnd(C, MAT_FINAL_ASSEMBLY)); 1189 PetscCall(MatSetOption(C, MAT_NEW_NONZERO_LOCATION_ERR, PETSC_TRUE)); 1190 1191 C->ops->matmultnumeric = MatMatMultNumeric_MPIAIJ_MPIAIJ_nonscalable; 1192 C->ops->productnumeric = MatProductNumeric_AB; 1193 1194 /* attach the supporting struct to C for reuse */ 1195 C->product->data = ptap; 1196 C->product->destroy = MatDestroy_MPIAIJ_MatMatMult; 1197 1198 /* set MatInfo */ 1199 afill = (PetscReal)api[am] / (adi[am] + aoi[am] + pi_loc[pm] + 1) + 1.e-5; 1200 if (afill < 1.0) afill = 1.0; 1201 C->info.mallocs = nspacedouble; 1202 C->info.fill_ratio_given = fill; 1203 C->info.fill_ratio_needed = afill; 1204 1205 #if defined(PETSC_USE_INFO) 1206 if (api[am]) { 1207 PetscCall(PetscInfo(C, "Reallocs %" PetscInt_FMT "; Fill ratio: given %g needed %g.\n", nspacedouble, (double)fill, (double)afill)); 1208 PetscCall(PetscInfo(C, "Use MatMatMult(A,B,MatReuse,%g,&C) for best performance.;\n", (double)afill)); 1209 } else { 1210 PetscCall(PetscInfo(C, "Empty matrix product\n")); 1211 } 1212 #endif 1213 1214 PetscCall(MatDestroy(&aopoth)); 1215 PetscCall(MatDestroy(&adpd)); 1216 PetscCall(PetscFree(j_temp)); 1217 PetscCall(PetscFree(adpoj)); 1218 PetscCall(PetscFree(adpoi)); 1219 PetscFunctionReturn(PETSC_SUCCESS); 1220 } 1221 1222 /* This routine only works when scall=MAT_REUSE_MATRIX! */ 1223 PetscErrorCode MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ_matmatmult(Mat P, Mat A, Mat C) 1224 { 1225 Mat_APMPI *ptap; 1226 Mat Pt; 1227 1228 PetscFunctionBegin; 1229 MatCheckProduct(C, 3); 1230 ptap = (Mat_APMPI *)C->product->data; 1231 PetscCheck(ptap, PetscObjectComm((PetscObject)C), PETSC_ERR_ARG_WRONGSTATE, "PtAP cannot be computed. Missing data"); 1232 PetscCheck(ptap->Pt, PetscObjectComm((PetscObject)C), PETSC_ERR_ARG_WRONGSTATE, "PtA cannot be reused. Do not call MatProductClear()"); 1233 1234 Pt = ptap->Pt; 1235 PetscCall(MatTransposeSetPrecursor(P, Pt)); 1236 PetscCall(MatTranspose(P, MAT_REUSE_MATRIX, &Pt)); 1237 PetscCall(MatMatMultNumeric_MPIAIJ_MPIAIJ(Pt, A, C)); 1238 PetscFunctionReturn(PETSC_SUCCESS); 1239 } 1240 1241 /* This routine is modified from MatPtAPSymbolic_MPIAIJ_MPIAIJ() */ 1242 PetscErrorCode MatTransposeMatMultSymbolic_MPIAIJ_MPIAIJ_nonscalable(Mat P, Mat A, PetscReal fill, Mat C) 1243 { 1244 Mat_APMPI *ptap; 1245 Mat_MPIAIJ *p = (Mat_MPIAIJ *)P->data; 1246 MPI_Comm comm; 1247 PetscMPIInt size, rank; 1248 PetscFreeSpaceList free_space = NULL, current_space = NULL; 1249 PetscInt pn = P->cmap->n, aN = A->cmap->N, an = A->cmap->n; 1250 PetscInt *lnk, i, k, rstart; 1251 PetscBT lnkbt; 1252 PetscMPIInt tagi, tagj, *len_si, *len_s, *len_ri, nrecv, proc, nsend; 1253 PETSC_UNUSED PetscMPIInt icompleted = 0; 1254 PetscInt **buf_rj, **buf_ri, **buf_ri_k, row, ncols, *cols; 1255 PetscInt len, *dnz, *onz, *owners, nzi; 1256 PetscInt nrows, *buf_s, *buf_si, *buf_si_i, **nextrow, **nextci; 1257 MPI_Request *swaits, *rwaits; 1258 MPI_Status *sstatus, rstatus; 1259 PetscLayout rowmap; 1260 PetscInt *owners_co, *coi, *coj; /* i and j array of (p->B)^T*A*P - used in the communication */ 1261 PetscMPIInt *len_r, *id_r; /* array of length of comm->size, store send/recv matrix values */ 1262 PetscInt *Jptr, *prmap = p->garray, con, j, Crmax; 1263 Mat_SeqAIJ *a_loc, *c_loc, *c_oth; 1264 PetscHMapI ta; 1265 MatType mtype; 1266 const char *prefix; 1267 1268 PetscFunctionBegin; 1269 PetscCall(PetscObjectGetComm((PetscObject)A, &comm)); 1270 PetscCallMPI(MPI_Comm_size(comm, &size)); 1271 PetscCallMPI(MPI_Comm_rank(comm, &rank)); 1272 1273 /* create symbolic parallel matrix C */ 1274 PetscCall(MatGetType(A, &mtype)); 1275 PetscCall(MatSetType(C, mtype)); 1276 1277 C->ops->transposematmultnumeric = MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ_nonscalable; 1278 1279 /* create struct Mat_APMPI and attached it to C later */ 1280 PetscCall(PetscNew(&ptap)); 1281 1282 /* (0) compute Rd = Pd^T, Ro = Po^T */ 1283 PetscCall(MatTranspose(p->A, MAT_INITIAL_MATRIX, &ptap->Rd)); 1284 PetscCall(MatTranspose(p->B, MAT_INITIAL_MATRIX, &ptap->Ro)); 1285 1286 /* (1) compute symbolic A_loc */ 1287 PetscCall(MatMPIAIJGetLocalMat(A, MAT_INITIAL_MATRIX, &ptap->A_loc)); 1288 1289 /* (2-1) compute symbolic C_oth = Ro*A_loc */ 1290 PetscCall(MatGetOptionsPrefix(A, &prefix)); 1291 PetscCall(MatSetOptionsPrefix(ptap->Ro, prefix)); 1292 PetscCall(MatAppendOptionsPrefix(ptap->Ro, "inner_offdiag_")); 1293 PetscCall(MatCreate(PETSC_COMM_SELF, &ptap->C_oth)); 1294 PetscCall(MatMatMultSymbolic_SeqAIJ_SeqAIJ(ptap->Ro, ptap->A_loc, fill, ptap->C_oth)); 1295 1296 /* (3) send coj of C_oth to other processors */ 1297 /* determine row ownership */ 1298 PetscCall(PetscLayoutCreate(comm, &rowmap)); 1299 rowmap->n = pn; 1300 rowmap->bs = 1; 1301 PetscCall(PetscLayoutSetUp(rowmap)); 1302 owners = rowmap->range; 1303 1304 /* determine the number of messages to send, their lengths */ 1305 PetscCall(PetscMalloc4(size, &len_s, size, &len_si, size, &sstatus, size + 1, &owners_co)); 1306 PetscCall(PetscArrayzero(len_s, size)); 1307 PetscCall(PetscArrayzero(len_si, size)); 1308 1309 c_oth = (Mat_SeqAIJ *)ptap->C_oth->data; 1310 coi = c_oth->i; 1311 coj = c_oth->j; 1312 con = ptap->C_oth->rmap->n; 1313 proc = 0; 1314 for (i = 0; i < con; i++) { 1315 while (prmap[i] >= owners[proc + 1]) proc++; 1316 len_si[proc]++; /* num of rows in Co(=Pt*A) to be sent to [proc] */ 1317 len_s[proc] += coi[i + 1] - coi[i]; /* num of nonzeros in Co to be sent to [proc] */ 1318 } 1319 1320 len = 0; /* max length of buf_si[], see (4) */ 1321 owners_co[0] = 0; 1322 nsend = 0; 1323 for (proc = 0; proc < size; proc++) { 1324 owners_co[proc + 1] = owners_co[proc] + len_si[proc]; 1325 if (len_s[proc]) { 1326 nsend++; 1327 len_si[proc] = 2 * (len_si[proc] + 1); /* length of buf_si to be sent to [proc] */ 1328 len += len_si[proc]; 1329 } 1330 } 1331 1332 /* determine the number and length of messages to receive for coi and coj */ 1333 PetscCall(PetscGatherNumberOfMessages(comm, NULL, len_s, &nrecv)); 1334 PetscCall(PetscGatherMessageLengths2(comm, nsend, nrecv, len_s, len_si, &id_r, &len_r, &len_ri)); 1335 1336 /* post the Irecv and Isend of coj */ 1337 PetscCall(PetscCommGetNewTag(comm, &tagj)); 1338 PetscCall(PetscPostIrecvInt(comm, tagj, nrecv, id_r, len_r, &buf_rj, &rwaits)); 1339 PetscCall(PetscMalloc1(nsend, &swaits)); 1340 for (proc = 0, k = 0; proc < size; proc++) { 1341 if (!len_s[proc]) continue; 1342 i = owners_co[proc]; 1343 PetscCallMPI(MPIU_Isend(coj + coi[i], len_s[proc], MPIU_INT, proc, tagj, comm, swaits + k)); 1344 k++; 1345 } 1346 1347 /* (2-2) compute symbolic C_loc = Rd*A_loc */ 1348 PetscCall(MatSetOptionsPrefix(ptap->Rd, prefix)); 1349 PetscCall(MatAppendOptionsPrefix(ptap->Rd, "inner_diag_")); 1350 PetscCall(MatCreate(PETSC_COMM_SELF, &ptap->C_loc)); 1351 PetscCall(MatMatMultSymbolic_SeqAIJ_SeqAIJ(ptap->Rd, ptap->A_loc, fill, ptap->C_loc)); 1352 c_loc = (Mat_SeqAIJ *)ptap->C_loc->data; 1353 1354 /* receives coj are complete */ 1355 for (i = 0; i < nrecv; i++) PetscCallMPI(MPI_Waitany(nrecv, rwaits, &icompleted, &rstatus)); 1356 PetscCall(PetscFree(rwaits)); 1357 if (nsend) PetscCallMPI(MPI_Waitall(nsend, swaits, sstatus)); 1358 1359 /* add received column indices into ta to update Crmax */ 1360 a_loc = (Mat_SeqAIJ *)ptap->A_loc->data; 1361 1362 /* create and initialize a linked list */ 1363 PetscCall(PetscHMapICreateWithSize(an, &ta)); /* for compute Crmax */ 1364 MatRowMergeMax_SeqAIJ(a_loc, ptap->A_loc->rmap->N, ta); 1365 1366 for (k = 0; k < nrecv; k++) { /* k-th received message */ 1367 Jptr = buf_rj[k]; 1368 for (j = 0; j < len_r[k]; j++) PetscCall(PetscHMapISet(ta, *(Jptr + j) + 1, 1)); 1369 } 1370 PetscCall(PetscHMapIGetSize(ta, &Crmax)); 1371 PetscCall(PetscHMapIDestroy(&ta)); 1372 1373 /* (4) send and recv coi */ 1374 PetscCall(PetscCommGetNewTag(comm, &tagi)); 1375 PetscCall(PetscPostIrecvInt(comm, tagi, nrecv, id_r, len_ri, &buf_ri, &rwaits)); 1376 PetscCall(PetscMalloc1(len, &buf_s)); 1377 buf_si = buf_s; /* points to the beginning of k-th msg to be sent */ 1378 for (proc = 0, k = 0; proc < size; proc++) { 1379 if (!len_s[proc]) continue; 1380 /* form outgoing message for i-structure: 1381 buf_si[0]: nrows to be sent 1382 [1:nrows]: row index (global) 1383 [nrows+1:2*nrows+1]: i-structure index 1384 */ 1385 nrows = len_si[proc] / 2 - 1; /* num of rows in Co to be sent to [proc] */ 1386 buf_si_i = buf_si + nrows + 1; 1387 buf_si[0] = nrows; 1388 buf_si_i[0] = 0; 1389 nrows = 0; 1390 for (i = owners_co[proc]; i < owners_co[proc + 1]; i++) { 1391 nzi = coi[i + 1] - coi[i]; 1392 buf_si_i[nrows + 1] = buf_si_i[nrows] + nzi; /* i-structure */ 1393 buf_si[nrows + 1] = prmap[i] - owners[proc]; /* local row index */ 1394 nrows++; 1395 } 1396 PetscCallMPI(MPIU_Isend(buf_si, len_si[proc], MPIU_INT, proc, tagi, comm, swaits + k)); 1397 k++; 1398 buf_si += len_si[proc]; 1399 } 1400 for (i = 0; i < nrecv; i++) PetscCallMPI(MPI_Waitany(nrecv, rwaits, &icompleted, &rstatus)); 1401 PetscCall(PetscFree(rwaits)); 1402 if (nsend) PetscCallMPI(MPI_Waitall(nsend, swaits, sstatus)); 1403 1404 PetscCall(PetscFree4(len_s, len_si, sstatus, owners_co)); 1405 PetscCall(PetscFree(len_ri)); 1406 PetscCall(PetscFree(swaits)); 1407 PetscCall(PetscFree(buf_s)); 1408 1409 /* (5) compute the local portion of C */ 1410 /* set initial free space to be Crmax, sufficient for holding nonzeros in each row of C */ 1411 PetscCall(PetscFreeSpaceGet(Crmax, &free_space)); 1412 current_space = free_space; 1413 1414 PetscCall(PetscMalloc3(nrecv, &buf_ri_k, nrecv, &nextrow, nrecv, &nextci)); 1415 for (k = 0; k < nrecv; k++) { 1416 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 1417 nrows = *buf_ri_k[k]; 1418 nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */ 1419 nextci[k] = buf_ri_k[k] + (nrows + 1); /* points to the next i-structure of k-th recved i-structure */ 1420 } 1421 1422 MatPreallocateBegin(comm, pn, an, dnz, onz); 1423 PetscCall(PetscLLCondensedCreate(Crmax, aN, &lnk, &lnkbt)); 1424 for (i = 0; i < pn; i++) { /* for each local row of C */ 1425 /* add C_loc into C */ 1426 nzi = c_loc->i[i + 1] - c_loc->i[i]; 1427 Jptr = c_loc->j + c_loc->i[i]; 1428 PetscCall(PetscLLCondensedAddSorted(nzi, Jptr, lnk, lnkbt)); 1429 1430 /* add received col data into lnk */ 1431 for (k = 0; k < nrecv; k++) { /* k-th received message */ 1432 if (i == *nextrow[k]) { /* i-th row */ 1433 nzi = *(nextci[k] + 1) - *nextci[k]; 1434 Jptr = buf_rj[k] + *nextci[k]; 1435 PetscCall(PetscLLCondensedAddSorted(nzi, Jptr, lnk, lnkbt)); 1436 nextrow[k]++; 1437 nextci[k]++; 1438 } 1439 } 1440 1441 /* add missing diagonal entry */ 1442 if (C->force_diagonals) { 1443 k = i + owners[rank]; /* column index */ 1444 PetscCall(PetscLLCondensedAddSorted(1, &k, lnk, lnkbt)); 1445 } 1446 1447 nzi = lnk[0]; 1448 1449 /* copy data into free space, then initialize lnk */ 1450 PetscCall(PetscLLCondensedClean(aN, nzi, current_space->array, lnk, lnkbt)); 1451 PetscCall(MatPreallocateSet(i + owners[rank], nzi, current_space->array, dnz, onz)); 1452 } 1453 PetscCall(PetscFree3(buf_ri_k, nextrow, nextci)); 1454 PetscCall(PetscLLDestroy(lnk, lnkbt)); 1455 PetscCall(PetscFreeSpaceDestroy(free_space)); 1456 1457 /* local sizes and preallocation */ 1458 PetscCall(MatSetSizes(C, pn, an, PETSC_DETERMINE, PETSC_DETERMINE)); 1459 PetscCall(PetscLayoutSetBlockSize(C->rmap, P->cmap->bs)); 1460 PetscCall(PetscLayoutSetBlockSize(C->cmap, A->cmap->bs)); 1461 PetscCall(MatMPIAIJSetPreallocation(C, 0, dnz, 0, onz)); 1462 MatPreallocateEnd(dnz, onz); 1463 1464 /* add C_loc and C_oth to C */ 1465 PetscCall(MatGetOwnershipRange(C, &rstart, NULL)); 1466 for (i = 0; i < pn; i++) { 1467 ncols = c_loc->i[i + 1] - c_loc->i[i]; 1468 cols = c_loc->j + c_loc->i[i]; 1469 row = rstart + i; 1470 PetscCall(MatSetValues(C, 1, (const PetscInt *)&row, ncols, (const PetscInt *)cols, NULL, INSERT_VALUES)); 1471 1472 if (C->force_diagonals) PetscCall(MatSetValues(C, 1, (const PetscInt *)&row, 1, (const PetscInt *)&row, NULL, INSERT_VALUES)); 1473 } 1474 for (i = 0; i < con; i++) { 1475 ncols = c_oth->i[i + 1] - c_oth->i[i]; 1476 cols = c_oth->j + c_oth->i[i]; 1477 row = prmap[i]; 1478 PetscCall(MatSetValues(C, 1, (const PetscInt *)&row, ncols, (const PetscInt *)cols, NULL, INSERT_VALUES)); 1479 } 1480 PetscCall(MatAssemblyBegin(C, MAT_FINAL_ASSEMBLY)); 1481 PetscCall(MatAssemblyEnd(C, MAT_FINAL_ASSEMBLY)); 1482 PetscCall(MatSetOption(C, MAT_NEW_NONZERO_LOCATION_ERR, PETSC_TRUE)); 1483 1484 /* members in merge */ 1485 PetscCall(PetscFree(id_r)); 1486 PetscCall(PetscFree(len_r)); 1487 PetscCall(PetscFree(buf_ri[0])); 1488 PetscCall(PetscFree(buf_ri)); 1489 PetscCall(PetscFree(buf_rj[0])); 1490 PetscCall(PetscFree(buf_rj)); 1491 PetscCall(PetscLayoutDestroy(&rowmap)); 1492 1493 /* attach the supporting struct to C for reuse */ 1494 C->product->data = ptap; 1495 C->product->destroy = MatDestroy_MPIAIJ_PtAP; 1496 PetscFunctionReturn(PETSC_SUCCESS); 1497 } 1498 1499 PetscErrorCode MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ_nonscalable(Mat P, Mat A, Mat C) 1500 { 1501 Mat_MPIAIJ *p = (Mat_MPIAIJ *)P->data; 1502 Mat_SeqAIJ *c_seq; 1503 Mat_APMPI *ptap; 1504 Mat A_loc, C_loc, C_oth; 1505 PetscInt i, rstart, rend, cm, ncols, row; 1506 const PetscInt *cols; 1507 const PetscScalar *vals; 1508 1509 PetscFunctionBegin; 1510 MatCheckProduct(C, 3); 1511 ptap = (Mat_APMPI *)C->product->data; 1512 PetscCheck(ptap, PetscObjectComm((PetscObject)C), PETSC_ERR_ARG_WRONGSTATE, "PtAP cannot be computed. Missing data"); 1513 PetscCheck(ptap->A_loc, PetscObjectComm((PetscObject)C), PETSC_ERR_ARG_WRONGSTATE, "PtA cannot be reused. Do not call MatProductClear()"); 1514 PetscCall(MatZeroEntries(C)); 1515 1516 /* These matrices are obtained in MatTransposeMatMultSymbolic() */ 1517 /* 1) get R = Pd^T, Ro = Po^T */ 1518 PetscCall(MatTransposeSetPrecursor(p->A, ptap->Rd)); 1519 PetscCall(MatTranspose(p->A, MAT_REUSE_MATRIX, &ptap->Rd)); 1520 PetscCall(MatTransposeSetPrecursor(p->B, ptap->Ro)); 1521 PetscCall(MatTranspose(p->B, MAT_REUSE_MATRIX, &ptap->Ro)); 1522 1523 /* 2) compute numeric A_loc */ 1524 PetscCall(MatMPIAIJGetLocalMat(A, MAT_REUSE_MATRIX, &ptap->A_loc)); 1525 1526 /* 3) C_loc = Rd*A_loc, C_oth = Ro*A_loc */ 1527 A_loc = ptap->A_loc; 1528 PetscCall(ptap->C_loc->ops->matmultnumeric(ptap->Rd, A_loc, ptap->C_loc)); 1529 PetscCall(ptap->C_oth->ops->matmultnumeric(ptap->Ro, A_loc, ptap->C_oth)); 1530 C_loc = ptap->C_loc; 1531 C_oth = ptap->C_oth; 1532 1533 /* add C_loc and C_oth to C */ 1534 PetscCall(MatGetOwnershipRange(C, &rstart, &rend)); 1535 1536 /* C_loc -> C */ 1537 cm = C_loc->rmap->N; 1538 c_seq = (Mat_SeqAIJ *)C_loc->data; 1539 cols = c_seq->j; 1540 vals = c_seq->a; 1541 for (i = 0; i < cm; i++) { 1542 ncols = c_seq->i[i + 1] - c_seq->i[i]; 1543 row = rstart + i; 1544 PetscCall(MatSetValues(C, 1, &row, ncols, cols, vals, ADD_VALUES)); 1545 cols += ncols; 1546 vals += ncols; 1547 } 1548 1549 /* Co -> C, off-processor part */ 1550 cm = C_oth->rmap->N; 1551 c_seq = (Mat_SeqAIJ *)C_oth->data; 1552 cols = c_seq->j; 1553 vals = c_seq->a; 1554 for (i = 0; i < cm; i++) { 1555 ncols = c_seq->i[i + 1] - c_seq->i[i]; 1556 row = p->garray[i]; 1557 PetscCall(MatSetValues(C, 1, &row, ncols, cols, vals, ADD_VALUES)); 1558 cols += ncols; 1559 vals += ncols; 1560 } 1561 PetscCall(MatAssemblyBegin(C, MAT_FINAL_ASSEMBLY)); 1562 PetscCall(MatAssemblyEnd(C, MAT_FINAL_ASSEMBLY)); 1563 PetscCall(MatSetOption(C, MAT_NEW_NONZERO_LOCATION_ERR, PETSC_TRUE)); 1564 PetscFunctionReturn(PETSC_SUCCESS); 1565 } 1566 1567 PetscErrorCode MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ(Mat P, Mat A, Mat C) 1568 { 1569 Mat_Merge_SeqsToMPI *merge; 1570 Mat_MPIAIJ *p = (Mat_MPIAIJ *)P->data; 1571 Mat_SeqAIJ *pd = (Mat_SeqAIJ *)p->A->data, *po = (Mat_SeqAIJ *)p->B->data; 1572 Mat_APMPI *ap; 1573 PetscInt *adj; 1574 PetscInt i, j, k, anz, pnz, row, *cj, nexta; 1575 MatScalar *ada, *ca, valtmp; 1576 PetscInt am = A->rmap->n, cm = C->rmap->n, pon = (p->B)->cmap->n; 1577 MPI_Comm comm; 1578 PetscMPIInt size, rank, taga, *len_s, proc; 1579 PetscInt *owners, nrows, **buf_ri_k, **nextrow, **nextci; 1580 PetscInt **buf_ri, **buf_rj; 1581 PetscInt cnz = 0, *bj_i, *bi, *bj, bnz, nextcj; /* bi,bj,ba: local array of C(mpi mat) */ 1582 MPI_Request *s_waits, *r_waits; 1583 MPI_Status *status; 1584 MatScalar **abuf_r, *ba_i, *pA, *coa, *ba; 1585 const PetscScalar *dummy; 1586 PetscInt *ai, *aj, *coi, *coj, *poJ, *pdJ; 1587 Mat A_loc; 1588 Mat_SeqAIJ *a_loc; 1589 1590 PetscFunctionBegin; 1591 MatCheckProduct(C, 3); 1592 ap = (Mat_APMPI *)C->product->data; 1593 PetscCheck(ap, PetscObjectComm((PetscObject)C), PETSC_ERR_ARG_WRONGSTATE, "PtA cannot be computed. Missing data"); 1594 PetscCheck(ap->A_loc, PetscObjectComm((PetscObject)C), PETSC_ERR_ARG_WRONGSTATE, "PtA cannot be reused. Do not call MatProductClear()"); 1595 PetscCall(PetscObjectGetComm((PetscObject)C, &comm)); 1596 PetscCallMPI(MPI_Comm_size(comm, &size)); 1597 PetscCallMPI(MPI_Comm_rank(comm, &rank)); 1598 1599 merge = ap->merge; 1600 1601 /* 2) compute numeric C_seq = P_loc^T*A_loc */ 1602 /* get data from symbolic products */ 1603 coi = merge->coi; 1604 coj = merge->coj; 1605 PetscCall(PetscCalloc1(coi[pon], &coa)); 1606 bi = merge->bi; 1607 bj = merge->bj; 1608 owners = merge->rowmap->range; 1609 PetscCall(PetscCalloc1(bi[cm], &ba)); 1610 1611 /* get A_loc by taking all local rows of A */ 1612 A_loc = ap->A_loc; 1613 PetscCall(MatMPIAIJGetLocalMat(A, MAT_REUSE_MATRIX, &A_loc)); 1614 a_loc = (Mat_SeqAIJ *)A_loc->data; 1615 ai = a_loc->i; 1616 aj = a_loc->j; 1617 1618 /* trigger copy to CPU */ 1619 PetscCall(MatSeqAIJGetArrayRead(p->A, &dummy)); 1620 PetscCall(MatSeqAIJRestoreArrayRead(p->A, &dummy)); 1621 PetscCall(MatSeqAIJGetArrayRead(p->B, &dummy)); 1622 PetscCall(MatSeqAIJRestoreArrayRead(p->B, &dummy)); 1623 for (i = 0; i < am; i++) { 1624 anz = ai[i + 1] - ai[i]; 1625 adj = aj + ai[i]; 1626 ada = a_loc->a + ai[i]; 1627 1628 /* 2-b) Compute Cseq = P_loc[i,:]^T*A[i,:] using outer product */ 1629 /* put the value into Co=(p->B)^T*A (off-diagonal part, send to others) */ 1630 pnz = po->i[i + 1] - po->i[i]; 1631 poJ = po->j + po->i[i]; 1632 pA = po->a + po->i[i]; 1633 for (j = 0; j < pnz; j++) { 1634 row = poJ[j]; 1635 cj = coj + coi[row]; 1636 ca = coa + coi[row]; 1637 /* perform sparse axpy */ 1638 nexta = 0; 1639 valtmp = pA[j]; 1640 for (k = 0; nexta < anz; k++) { 1641 if (cj[k] == adj[nexta]) { 1642 ca[k] += valtmp * ada[nexta]; 1643 nexta++; 1644 } 1645 } 1646 PetscCall(PetscLogFlops(2.0 * anz)); 1647 } 1648 1649 /* put the value into Cd (diagonal part) */ 1650 pnz = pd->i[i + 1] - pd->i[i]; 1651 pdJ = pd->j + pd->i[i]; 1652 pA = pd->a + pd->i[i]; 1653 for (j = 0; j < pnz; j++) { 1654 row = pdJ[j]; 1655 cj = bj + bi[row]; 1656 ca = ba + bi[row]; 1657 /* perform sparse axpy */ 1658 nexta = 0; 1659 valtmp = pA[j]; 1660 for (k = 0; nexta < anz; k++) { 1661 if (cj[k] == adj[nexta]) { 1662 ca[k] += valtmp * ada[nexta]; 1663 nexta++; 1664 } 1665 } 1666 PetscCall(PetscLogFlops(2.0 * anz)); 1667 } 1668 } 1669 1670 /* 3) send and recv matrix values coa */ 1671 buf_ri = merge->buf_ri; 1672 buf_rj = merge->buf_rj; 1673 len_s = merge->len_s; 1674 PetscCall(PetscCommGetNewTag(comm, &taga)); 1675 PetscCall(PetscPostIrecvScalar(comm, taga, merge->nrecv, merge->id_r, merge->len_r, &abuf_r, &r_waits)); 1676 1677 PetscCall(PetscMalloc2(merge->nsend, &s_waits, size, &status)); 1678 for (proc = 0, k = 0; proc < size; proc++) { 1679 if (!len_s[proc]) continue; 1680 i = merge->owners_co[proc]; 1681 PetscCallMPI(MPIU_Isend(coa + coi[i], len_s[proc], MPIU_MATSCALAR, proc, taga, comm, s_waits + k)); 1682 k++; 1683 } 1684 if (merge->nrecv) PetscCallMPI(MPI_Waitall(merge->nrecv, r_waits, status)); 1685 if (merge->nsend) PetscCallMPI(MPI_Waitall(merge->nsend, s_waits, status)); 1686 1687 PetscCall(PetscFree2(s_waits, status)); 1688 PetscCall(PetscFree(r_waits)); 1689 PetscCall(PetscFree(coa)); 1690 1691 /* 4) insert local Cseq and received values into Cmpi */ 1692 PetscCall(PetscMalloc3(merge->nrecv, &buf_ri_k, merge->nrecv, &nextrow, merge->nrecv, &nextci)); 1693 for (k = 0; k < merge->nrecv; k++) { 1694 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 1695 nrows = *buf_ri_k[k]; 1696 nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */ 1697 nextci[k] = buf_ri_k[k] + (nrows + 1); /* points to the next i-structure of k-th recved i-structure */ 1698 } 1699 1700 for (i = 0; i < cm; i++) { 1701 row = owners[rank] + i; /* global row index of C_seq */ 1702 bj_i = bj + bi[i]; /* col indices of the i-th row of C */ 1703 ba_i = ba + bi[i]; 1704 bnz = bi[i + 1] - bi[i]; 1705 /* add received vals into ba */ 1706 for (k = 0; k < merge->nrecv; k++) { /* k-th received message */ 1707 /* i-th row */ 1708 if (i == *nextrow[k]) { 1709 cnz = *(nextci[k] + 1) - *nextci[k]; 1710 cj = buf_rj[k] + *nextci[k]; 1711 ca = abuf_r[k] + *nextci[k]; 1712 nextcj = 0; 1713 for (j = 0; nextcj < cnz; j++) { 1714 if (bj_i[j] == cj[nextcj]) { /* bcol == ccol */ 1715 ba_i[j] += ca[nextcj++]; 1716 } 1717 } 1718 nextrow[k]++; 1719 nextci[k]++; 1720 PetscCall(PetscLogFlops(2.0 * cnz)); 1721 } 1722 } 1723 PetscCall(MatSetValues(C, 1, &row, bnz, bj_i, ba_i, INSERT_VALUES)); 1724 } 1725 PetscCall(MatAssemblyBegin(C, MAT_FINAL_ASSEMBLY)); 1726 PetscCall(MatAssemblyEnd(C, MAT_FINAL_ASSEMBLY)); 1727 1728 PetscCall(PetscFree(ba)); 1729 PetscCall(PetscFree(abuf_r[0])); 1730 PetscCall(PetscFree(abuf_r)); 1731 PetscCall(PetscFree3(buf_ri_k, nextrow, nextci)); 1732 PetscFunctionReturn(PETSC_SUCCESS); 1733 } 1734 1735 PetscErrorCode MatTransposeMatMultSymbolic_MPIAIJ_MPIAIJ(Mat P, Mat A, PetscReal fill, Mat C) 1736 { 1737 Mat A_loc; 1738 Mat_APMPI *ap; 1739 PetscFreeSpaceList free_space = NULL, current_space = NULL; 1740 Mat_MPIAIJ *p = (Mat_MPIAIJ *)P->data, *a = (Mat_MPIAIJ *)A->data; 1741 PetscInt *pdti, *pdtj, *poti, *potj, *ptJ; 1742 PetscInt nnz; 1743 PetscInt *lnk, *owners_co, *coi, *coj, i, k, pnz, row; 1744 PetscInt am = A->rmap->n, pn = P->cmap->n; 1745 MPI_Comm comm; 1746 PetscMPIInt size, rank, tagi, tagj, *len_si, *len_s, *len_ri, proc; 1747 PetscInt **buf_rj, **buf_ri, **buf_ri_k; 1748 PetscInt len, *dnz, *onz, *owners; 1749 PetscInt nzi, *bi, *bj; 1750 PetscInt nrows, *buf_s, *buf_si, *buf_si_i, **nextrow, **nextci; 1751 MPI_Request *swaits, *rwaits; 1752 MPI_Status *sstatus, rstatus; 1753 Mat_Merge_SeqsToMPI *merge; 1754 PetscInt *ai, *aj, *Jptr, anz, *prmap = p->garray, pon, nspacedouble = 0, j; 1755 PetscReal afill = 1.0, afill_tmp; 1756 PetscInt rstart = P->cmap->rstart, rmax, Armax; 1757 Mat_SeqAIJ *a_loc; 1758 PetscHMapI ta; 1759 MatType mtype; 1760 1761 PetscFunctionBegin; 1762 PetscCall(PetscObjectGetComm((PetscObject)A, &comm)); 1763 /* check if matrix local sizes are compatible */ 1764 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, 1765 A->rmap->rend, P->rmap->rstart, P->rmap->rend); 1766 1767 PetscCallMPI(MPI_Comm_size(comm, &size)); 1768 PetscCallMPI(MPI_Comm_rank(comm, &rank)); 1769 1770 /* create struct Mat_APMPI and attached it to C later */ 1771 PetscCall(PetscNew(&ap)); 1772 1773 /* get A_loc by taking all local rows of A */ 1774 PetscCall(MatMPIAIJGetLocalMat(A, MAT_INITIAL_MATRIX, &A_loc)); 1775 1776 ap->A_loc = A_loc; 1777 a_loc = (Mat_SeqAIJ *)A_loc->data; 1778 ai = a_loc->i; 1779 aj = a_loc->j; 1780 1781 /* determine symbolic Co=(p->B)^T*A - send to others */ 1782 PetscCall(MatGetSymbolicTranspose_SeqAIJ(p->A, &pdti, &pdtj)); 1783 PetscCall(MatGetSymbolicTranspose_SeqAIJ(p->B, &poti, &potj)); 1784 pon = (p->B)->cmap->n; /* total num of rows to be sent to other processors 1785 >= (num of nonzero rows of C_seq) - pn */ 1786 PetscCall(PetscMalloc1(pon + 1, &coi)); 1787 coi[0] = 0; 1788 1789 /* set initial free space to be fill*(nnz(p->B) + nnz(A)) */ 1790 nnz = PetscRealIntMultTruncate(fill, PetscIntSumTruncate(poti[pon], ai[am])); 1791 PetscCall(PetscFreeSpaceGet(nnz, &free_space)); 1792 current_space = free_space; 1793 1794 /* create and initialize a linked list */ 1795 PetscCall(PetscHMapICreateWithSize(A->cmap->n + a->B->cmap->N, &ta)); 1796 MatRowMergeMax_SeqAIJ(a_loc, am, ta); 1797 PetscCall(PetscHMapIGetSize(ta, &Armax)); 1798 1799 PetscCall(PetscLLCondensedCreate_Scalable(Armax, &lnk)); 1800 1801 for (i = 0; i < pon; i++) { 1802 pnz = poti[i + 1] - poti[i]; 1803 ptJ = potj + poti[i]; 1804 for (j = 0; j < pnz; j++) { 1805 row = ptJ[j]; /* row of A_loc == col of Pot */ 1806 anz = ai[row + 1] - ai[row]; 1807 Jptr = aj + ai[row]; 1808 /* add non-zero cols of AP into the sorted linked list lnk */ 1809 PetscCall(PetscLLCondensedAddSorted_Scalable(anz, Jptr, lnk)); 1810 } 1811 nnz = lnk[0]; 1812 1813 /* If free space is not available, double the total space in the list */ 1814 if (current_space->local_remaining < nnz) { 1815 PetscCall(PetscFreeSpaceGet(PetscIntSumTruncate(nnz, current_space->total_array_size), ¤t_space)); 1816 nspacedouble++; 1817 } 1818 1819 /* Copy data into free space, and zero out denserows */ 1820 PetscCall(PetscLLCondensedClean_Scalable(nnz, current_space->array, lnk)); 1821 1822 current_space->array += nnz; 1823 current_space->local_used += nnz; 1824 current_space->local_remaining -= nnz; 1825 1826 coi[i + 1] = coi[i] + nnz; 1827 } 1828 1829 PetscCall(PetscMalloc1(coi[pon], &coj)); 1830 PetscCall(PetscFreeSpaceContiguous(&free_space, coj)); 1831 PetscCall(PetscLLCondensedDestroy_Scalable(lnk)); /* must destroy to get a new one for C */ 1832 1833 afill_tmp = (PetscReal)coi[pon] / (poti[pon] + ai[am] + 1); 1834 if (afill_tmp > afill) afill = afill_tmp; 1835 1836 /* send j-array (coj) of Co to other processors */ 1837 /* determine row ownership */ 1838 PetscCall(PetscNew(&merge)); 1839 PetscCall(PetscLayoutCreate(comm, &merge->rowmap)); 1840 1841 merge->rowmap->n = pn; 1842 merge->rowmap->bs = 1; 1843 1844 PetscCall(PetscLayoutSetUp(merge->rowmap)); 1845 owners = merge->rowmap->range; 1846 1847 /* determine the number of messages to send, their lengths */ 1848 PetscCall(PetscCalloc1(size, &len_si)); 1849 PetscCall(PetscCalloc1(size, &merge->len_s)); 1850 1851 len_s = merge->len_s; 1852 merge->nsend = 0; 1853 1854 PetscCall(PetscMalloc1(size + 1, &owners_co)); 1855 1856 proc = 0; 1857 for (i = 0; i < pon; i++) { 1858 while (prmap[i] >= owners[proc + 1]) proc++; 1859 len_si[proc]++; /* num of rows in Co to be sent to [proc] */ 1860 len_s[proc] += coi[i + 1] - coi[i]; 1861 } 1862 1863 len = 0; /* max length of buf_si[] */ 1864 owners_co[0] = 0; 1865 for (proc = 0; proc < size; proc++) { 1866 owners_co[proc + 1] = owners_co[proc] + len_si[proc]; 1867 if (len_s[proc]) { 1868 merge->nsend++; 1869 len_si[proc] = 2 * (len_si[proc] + 1); 1870 len += len_si[proc]; 1871 } 1872 } 1873 1874 /* determine the number and length of messages to receive for coi and coj */ 1875 PetscCall(PetscGatherNumberOfMessages(comm, NULL, len_s, &merge->nrecv)); 1876 PetscCall(PetscGatherMessageLengths2(comm, merge->nsend, merge->nrecv, len_s, len_si, &merge->id_r, &merge->len_r, &len_ri)); 1877 1878 /* post the Irecv and Isend of coj */ 1879 PetscCall(PetscCommGetNewTag(comm, &tagj)); 1880 PetscCall(PetscPostIrecvInt(comm, tagj, merge->nrecv, merge->id_r, merge->len_r, &buf_rj, &rwaits)); 1881 PetscCall(PetscMalloc1(merge->nsend, &swaits)); 1882 for (proc = 0, k = 0; proc < size; proc++) { 1883 if (!len_s[proc]) continue; 1884 i = owners_co[proc]; 1885 PetscCallMPI(MPIU_Isend(coj + coi[i], len_s[proc], MPIU_INT, proc, tagj, comm, swaits + k)); 1886 k++; 1887 } 1888 1889 /* receives and sends of coj are complete */ 1890 PetscCall(PetscMalloc1(size, &sstatus)); 1891 for (i = 0; i < merge->nrecv; i++) { 1892 PETSC_UNUSED PetscMPIInt icompleted; 1893 PetscCallMPI(MPI_Waitany(merge->nrecv, rwaits, &icompleted, &rstatus)); 1894 } 1895 PetscCall(PetscFree(rwaits)); 1896 if (merge->nsend) PetscCallMPI(MPI_Waitall(merge->nsend, swaits, sstatus)); 1897 1898 /* add received column indices into table to update Armax */ 1899 /* Armax can be as large as aN if a P[row,:] is dense, see src/ksp/ksp/tutorials/ex56.c! */ 1900 for (k = 0; k < merge->nrecv; k++) { /* k-th received message */ 1901 Jptr = buf_rj[k]; 1902 for (j = 0; j < merge->len_r[k]; j++) PetscCall(PetscHMapISet(ta, *(Jptr + j) + 1, 1)); 1903 } 1904 PetscCall(PetscHMapIGetSize(ta, &Armax)); 1905 1906 /* send and recv coi */ 1907 PetscCall(PetscCommGetNewTag(comm, &tagi)); 1908 PetscCall(PetscPostIrecvInt(comm, tagi, merge->nrecv, merge->id_r, len_ri, &buf_ri, &rwaits)); 1909 PetscCall(PetscMalloc1(len, &buf_s)); 1910 buf_si = buf_s; /* points to the beginning of k-th msg to be sent */ 1911 for (proc = 0, k = 0; proc < size; proc++) { 1912 if (!len_s[proc]) continue; 1913 /* form outgoing message for i-structure: 1914 buf_si[0]: nrows to be sent 1915 [1:nrows]: row index (global) 1916 [nrows+1:2*nrows+1]: i-structure index 1917 */ 1918 nrows = len_si[proc] / 2 - 1; 1919 buf_si_i = buf_si + nrows + 1; 1920 buf_si[0] = nrows; 1921 buf_si_i[0] = 0; 1922 nrows = 0; 1923 for (i = owners_co[proc]; i < owners_co[proc + 1]; i++) { 1924 nzi = coi[i + 1] - coi[i]; 1925 buf_si_i[nrows + 1] = buf_si_i[nrows] + nzi; /* i-structure */ 1926 buf_si[nrows + 1] = prmap[i] - owners[proc]; /* local row index */ 1927 nrows++; 1928 } 1929 PetscCallMPI(MPIU_Isend(buf_si, len_si[proc], MPIU_INT, proc, tagi, comm, swaits + k)); 1930 k++; 1931 buf_si += len_si[proc]; 1932 } 1933 i = merge->nrecv; 1934 while (i--) { 1935 PETSC_UNUSED PetscMPIInt icompleted; 1936 PetscCallMPI(MPI_Waitany(merge->nrecv, rwaits, &icompleted, &rstatus)); 1937 } 1938 PetscCall(PetscFree(rwaits)); 1939 if (merge->nsend) PetscCallMPI(MPI_Waitall(merge->nsend, swaits, sstatus)); 1940 PetscCall(PetscFree(len_si)); 1941 PetscCall(PetscFree(len_ri)); 1942 PetscCall(PetscFree(swaits)); 1943 PetscCall(PetscFree(sstatus)); 1944 PetscCall(PetscFree(buf_s)); 1945 1946 /* compute the local portion of C (mpi mat) */ 1947 /* allocate bi array and free space for accumulating nonzero column info */ 1948 PetscCall(PetscMalloc1(pn + 1, &bi)); 1949 bi[0] = 0; 1950 1951 /* set initial free space to be fill*(nnz(P) + nnz(AP)) */ 1952 nnz = PetscRealIntMultTruncate(fill, PetscIntSumTruncate(pdti[pn], PetscIntSumTruncate(poti[pon], ai[am]))); 1953 PetscCall(PetscFreeSpaceGet(nnz, &free_space)); 1954 current_space = free_space; 1955 1956 PetscCall(PetscMalloc3(merge->nrecv, &buf_ri_k, merge->nrecv, &nextrow, merge->nrecv, &nextci)); 1957 for (k = 0; k < merge->nrecv; k++) { 1958 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 1959 nrows = *buf_ri_k[k]; 1960 nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */ 1961 nextci[k] = buf_ri_k[k] + (nrows + 1); /* points to the next i-structure of k-th received i-structure */ 1962 } 1963 1964 PetscCall(PetscLLCondensedCreate_Scalable(Armax, &lnk)); 1965 MatPreallocateBegin(comm, pn, A->cmap->n, dnz, onz); 1966 rmax = 0; 1967 for (i = 0; i < pn; i++) { 1968 /* add pdt[i,:]*AP into lnk */ 1969 pnz = pdti[i + 1] - pdti[i]; 1970 ptJ = pdtj + pdti[i]; 1971 for (j = 0; j < pnz; j++) { 1972 row = ptJ[j]; /* row of AP == col of Pt */ 1973 anz = ai[row + 1] - ai[row]; 1974 Jptr = aj + ai[row]; 1975 /* add non-zero cols of AP into the sorted linked list lnk */ 1976 PetscCall(PetscLLCondensedAddSorted_Scalable(anz, Jptr, lnk)); 1977 } 1978 1979 /* add received col data into lnk */ 1980 for (k = 0; k < merge->nrecv; k++) { /* k-th received message */ 1981 if (i == *nextrow[k]) { /* i-th row */ 1982 nzi = *(nextci[k] + 1) - *nextci[k]; 1983 Jptr = buf_rj[k] + *nextci[k]; 1984 PetscCall(PetscLLCondensedAddSorted_Scalable(nzi, Jptr, lnk)); 1985 nextrow[k]++; 1986 nextci[k]++; 1987 } 1988 } 1989 1990 /* add missing diagonal entry */ 1991 if (C->force_diagonals) { 1992 k = i + owners[rank]; /* column index */ 1993 PetscCall(PetscLLCondensedAddSorted_Scalable(1, &k, lnk)); 1994 } 1995 1996 nnz = lnk[0]; 1997 1998 /* if free space is not available, make more free space */ 1999 if (current_space->local_remaining < nnz) { 2000 PetscCall(PetscFreeSpaceGet(PetscIntSumTruncate(nnz, current_space->total_array_size), ¤t_space)); 2001 nspacedouble++; 2002 } 2003 /* copy data into free space, then initialize lnk */ 2004 PetscCall(PetscLLCondensedClean_Scalable(nnz, current_space->array, lnk)); 2005 PetscCall(MatPreallocateSet(i + owners[rank], nnz, current_space->array, dnz, onz)); 2006 2007 current_space->array += nnz; 2008 current_space->local_used += nnz; 2009 current_space->local_remaining -= nnz; 2010 2011 bi[i + 1] = bi[i] + nnz; 2012 if (nnz > rmax) rmax = nnz; 2013 } 2014 PetscCall(PetscFree3(buf_ri_k, nextrow, nextci)); 2015 2016 PetscCall(PetscMalloc1(bi[pn], &bj)); 2017 PetscCall(PetscFreeSpaceContiguous(&free_space, bj)); 2018 afill_tmp = (PetscReal)bi[pn] / (pdti[pn] + poti[pon] + ai[am] + 1); 2019 if (afill_tmp > afill) afill = afill_tmp; 2020 PetscCall(PetscLLCondensedDestroy_Scalable(lnk)); 2021 PetscCall(PetscHMapIDestroy(&ta)); 2022 PetscCall(MatRestoreSymbolicTranspose_SeqAIJ(p->A, &pdti, &pdtj)); 2023 PetscCall(MatRestoreSymbolicTranspose_SeqAIJ(p->B, &poti, &potj)); 2024 2025 /* create symbolic parallel matrix C - why cannot be assembled in Numeric part */ 2026 PetscCall(MatSetSizes(C, pn, A->cmap->n, PETSC_DETERMINE, PETSC_DETERMINE)); 2027 PetscCall(MatSetBlockSizes(C, P->cmap->bs, A->cmap->bs)); 2028 PetscCall(MatGetType(A, &mtype)); 2029 PetscCall(MatSetType(C, mtype)); 2030 PetscCall(MatMPIAIJSetPreallocation(C, 0, dnz, 0, onz)); 2031 MatPreallocateEnd(dnz, onz); 2032 PetscCall(MatSetBlockSize(C, 1)); 2033 PetscCall(MatSetOption(C, MAT_NO_OFF_PROC_ENTRIES, PETSC_TRUE)); 2034 for (i = 0; i < pn; i++) { 2035 row = i + rstart; 2036 nnz = bi[i + 1] - bi[i]; 2037 Jptr = bj + bi[i]; 2038 PetscCall(MatSetValues(C, 1, &row, nnz, Jptr, NULL, INSERT_VALUES)); 2039 } 2040 PetscCall(MatAssemblyBegin(C, MAT_FINAL_ASSEMBLY)); 2041 PetscCall(MatAssemblyEnd(C, MAT_FINAL_ASSEMBLY)); 2042 PetscCall(MatSetOption(C, MAT_NEW_NONZERO_LOCATION_ERR, PETSC_TRUE)); 2043 merge->bi = bi; 2044 merge->bj = bj; 2045 merge->coi = coi; 2046 merge->coj = coj; 2047 merge->buf_ri = buf_ri; 2048 merge->buf_rj = buf_rj; 2049 merge->owners_co = owners_co; 2050 2051 /* attach the supporting struct to C for reuse */ 2052 C->product->data = ap; 2053 C->product->destroy = MatDestroy_MPIAIJ_PtAP; 2054 ap->merge = merge; 2055 2056 C->ops->mattransposemultnumeric = MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ; 2057 2058 #if defined(PETSC_USE_INFO) 2059 if (bi[pn] != 0) { 2060 PetscCall(PetscInfo(C, "Reallocs %" PetscInt_FMT "; Fill ratio: given %g needed %g.\n", nspacedouble, (double)fill, (double)afill)); 2061 PetscCall(PetscInfo(C, "Use MatTransposeMatMult(A,B,MatReuse,%g,&C) for best performance.\n", (double)afill)); 2062 } else { 2063 PetscCall(PetscInfo(C, "Empty matrix product\n")); 2064 } 2065 #endif 2066 PetscFunctionReturn(PETSC_SUCCESS); 2067 } 2068 2069 static PetscErrorCode MatProductSymbolic_AtB_MPIAIJ_MPIAIJ(Mat C) 2070 { 2071 Mat_Product *product = C->product; 2072 Mat A = product->A, B = product->B; 2073 PetscReal fill = product->fill; 2074 PetscBool flg; 2075 2076 PetscFunctionBegin; 2077 /* scalable */ 2078 PetscCall(PetscStrcmp(product->alg, "scalable", &flg)); 2079 if (flg) { 2080 PetscCall(MatTransposeMatMultSymbolic_MPIAIJ_MPIAIJ(A, B, fill, C)); 2081 goto next; 2082 } 2083 2084 /* nonscalable */ 2085 PetscCall(PetscStrcmp(product->alg, "nonscalable", &flg)); 2086 if (flg) { 2087 PetscCall(MatTransposeMatMultSymbolic_MPIAIJ_MPIAIJ_nonscalable(A, B, fill, C)); 2088 goto next; 2089 } 2090 2091 /* matmatmult */ 2092 PetscCall(PetscStrcmp(product->alg, "at*b", &flg)); 2093 if (flg) { 2094 Mat At; 2095 Mat_APMPI *ptap; 2096 2097 PetscCall(MatTranspose(A, MAT_INITIAL_MATRIX, &At)); 2098 PetscCall(MatMatMultSymbolic_MPIAIJ_MPIAIJ(At, B, fill, C)); 2099 ptap = (Mat_APMPI *)C->product->data; 2100 if (ptap) { 2101 ptap->Pt = At; 2102 C->product->destroy = MatDestroy_MPIAIJ_PtAP; 2103 } 2104 C->ops->transposematmultnumeric = MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ_matmatmult; 2105 goto next; 2106 } 2107 2108 /* backend general code */ 2109 PetscCall(PetscStrcmp(product->alg, "backend", &flg)); 2110 if (flg) { 2111 PetscCall(MatProductSymbolic_MPIAIJBACKEND(C)); 2112 PetscFunctionReturn(PETSC_SUCCESS); 2113 } 2114 2115 SETERRQ(PETSC_COMM_SELF, PETSC_ERR_SUP, "MatProduct type is not supported"); 2116 2117 next: 2118 C->ops->productnumeric = MatProductNumeric_AtB; 2119 PetscFunctionReturn(PETSC_SUCCESS); 2120 } 2121 2122 /* Set options for MatMatMultxxx_MPIAIJ_MPIAIJ */ 2123 static PetscErrorCode MatProductSetFromOptions_MPIAIJ_AB(Mat C) 2124 { 2125 Mat_Product *product = C->product; 2126 Mat A = product->A, B = product->B; 2127 #if defined(PETSC_HAVE_HYPRE) 2128 const char *algTypes[5] = {"scalable", "nonscalable", "seqmpi", "backend", "hypre"}; 2129 PetscInt nalg = 5; 2130 #else 2131 const char *algTypes[4] = { 2132 "scalable", 2133 "nonscalable", 2134 "seqmpi", 2135 "backend", 2136 }; 2137 PetscInt nalg = 4; 2138 #endif 2139 PetscInt alg = 1; /* set nonscalable algorithm as default */ 2140 PetscBool flg; 2141 MPI_Comm comm; 2142 2143 PetscFunctionBegin; 2144 PetscCall(PetscObjectGetComm((PetscObject)C, &comm)); 2145 2146 /* Set "nonscalable" as default algorithm */ 2147 PetscCall(PetscStrcmp(C->product->alg, "default", &flg)); 2148 if (flg) { 2149 PetscCall(MatProductSetAlgorithm(C, algTypes[alg])); 2150 2151 /* Set "scalable" as default if BN and local nonzeros of A and B are large */ 2152 if (B->cmap->N > 100000) { /* may switch to scalable algorithm as default */ 2153 MatInfo Ainfo, Binfo; 2154 PetscInt nz_local; 2155 PetscBool alg_scalable_loc = PETSC_FALSE, alg_scalable; 2156 2157 PetscCall(MatGetInfo(A, MAT_LOCAL, &Ainfo)); 2158 PetscCall(MatGetInfo(B, MAT_LOCAL, &Binfo)); 2159 nz_local = (PetscInt)(Ainfo.nz_allocated + Binfo.nz_allocated); 2160 2161 if (B->cmap->N > product->fill * nz_local) alg_scalable_loc = PETSC_TRUE; 2162 PetscCallMPI(MPIU_Allreduce(&alg_scalable_loc, &alg_scalable, 1, MPI_C_BOOL, MPI_LOR, comm)); 2163 2164 if (alg_scalable) { 2165 alg = 0; /* scalable algorithm would 50% slower than nonscalable algorithm */ 2166 PetscCall(MatProductSetAlgorithm(C, algTypes[alg])); 2167 PetscCall(PetscInfo(B, "Use scalable algorithm, BN %" PetscInt_FMT ", fill*nz_allocated %g\n", B->cmap->N, (double)(product->fill * nz_local))); 2168 } 2169 } 2170 } 2171 2172 /* Get runtime option */ 2173 if (product->api_user) { 2174 PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatMatMult", "Mat"); 2175 PetscCall(PetscOptionsEList("-matmatmult_via", "Algorithmic approach", "MatMatMult", algTypes, nalg, algTypes[alg], &alg, &flg)); 2176 PetscOptionsEnd(); 2177 } else { 2178 PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatProduct_AB", "Mat"); 2179 PetscCall(PetscOptionsEList("-mat_product_algorithm", "Algorithmic approach", "MatMatMult", algTypes, nalg, algTypes[alg], &alg, &flg)); 2180 PetscOptionsEnd(); 2181 } 2182 if (flg) PetscCall(MatProductSetAlgorithm(C, algTypes[alg])); 2183 2184 C->ops->productsymbolic = MatProductSymbolic_AB_MPIAIJ_MPIAIJ; 2185 PetscFunctionReturn(PETSC_SUCCESS); 2186 } 2187 2188 static PetscErrorCode MatProductSetFromOptions_MPIAIJ_ABt(Mat C) 2189 { 2190 PetscFunctionBegin; 2191 PetscCall(MatProductSetFromOptions_MPIAIJ_AB(C)); 2192 C->ops->productsymbolic = MatProductSymbolic_ABt_MPIAIJ_MPIAIJ; 2193 PetscFunctionReturn(PETSC_SUCCESS); 2194 } 2195 2196 /* Set options for MatTransposeMatMultXXX_MPIAIJ_MPIAIJ */ 2197 static PetscErrorCode MatProductSetFromOptions_MPIAIJ_AtB(Mat C) 2198 { 2199 Mat_Product *product = C->product; 2200 Mat A = product->A, B = product->B; 2201 const char *algTypes[4] = {"scalable", "nonscalable", "at*b", "backend"}; 2202 PetscInt nalg = 4; 2203 PetscInt alg = 1; /* set default algorithm */ 2204 PetscBool flg; 2205 MPI_Comm comm; 2206 2207 PetscFunctionBegin; 2208 /* Check matrix local sizes */ 2209 PetscCall(PetscObjectGetComm((PetscObject)C, &comm)); 2210 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 ")", 2211 A->rmap->rstart, A->rmap->rend, B->rmap->rstart, B->rmap->rend); 2212 2213 /* Set default algorithm */ 2214 PetscCall(PetscStrcmp(C->product->alg, "default", &flg)); 2215 if (flg) PetscCall(MatProductSetAlgorithm(C, algTypes[alg])); 2216 2217 /* Set "scalable" as default if BN and local nonzeros of A and B are large */ 2218 if (alg && B->cmap->N > 100000) { /* may switch to scalable algorithm as default */ 2219 MatInfo Ainfo, Binfo; 2220 PetscInt nz_local; 2221 PetscBool alg_scalable_loc = PETSC_FALSE, alg_scalable; 2222 2223 PetscCall(MatGetInfo(A, MAT_LOCAL, &Ainfo)); 2224 PetscCall(MatGetInfo(B, MAT_LOCAL, &Binfo)); 2225 nz_local = (PetscInt)(Ainfo.nz_allocated + Binfo.nz_allocated); 2226 2227 if (B->cmap->N > product->fill * nz_local) alg_scalable_loc = PETSC_TRUE; 2228 PetscCallMPI(MPIU_Allreduce(&alg_scalable_loc, &alg_scalable, 1, MPI_C_BOOL, MPI_LOR, comm)); 2229 2230 if (alg_scalable) { 2231 alg = 0; /* scalable algorithm would 50% slower than nonscalable algorithm */ 2232 PetscCall(MatProductSetAlgorithm(C, algTypes[alg])); 2233 PetscCall(PetscInfo(B, "Use scalable algorithm, BN %" PetscInt_FMT ", fill*nz_allocated %g\n", B->cmap->N, (double)(product->fill * nz_local))); 2234 } 2235 } 2236 2237 /* Get runtime option */ 2238 if (product->api_user) { 2239 PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatTransposeMatMult", "Mat"); 2240 PetscCall(PetscOptionsEList("-mattransposematmult_via", "Algorithmic approach", "MatTransposeMatMult", algTypes, nalg, algTypes[alg], &alg, &flg)); 2241 PetscOptionsEnd(); 2242 } else { 2243 PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatProduct_AtB", "Mat"); 2244 PetscCall(PetscOptionsEList("-mat_product_algorithm", "Algorithmic approach", "MatTransposeMatMult", algTypes, nalg, algTypes[alg], &alg, &flg)); 2245 PetscOptionsEnd(); 2246 } 2247 if (flg) PetscCall(MatProductSetAlgorithm(C, algTypes[alg])); 2248 2249 C->ops->productsymbolic = MatProductSymbolic_AtB_MPIAIJ_MPIAIJ; 2250 PetscFunctionReturn(PETSC_SUCCESS); 2251 } 2252 2253 static PetscErrorCode MatProductSetFromOptions_MPIAIJ_PtAP(Mat C) 2254 { 2255 Mat_Product *product = C->product; 2256 Mat A = product->A, P = product->B; 2257 MPI_Comm comm; 2258 PetscBool flg; 2259 PetscInt alg = 1; /* set default algorithm */ 2260 #if !defined(PETSC_HAVE_HYPRE) 2261 const char *algTypes[5] = {"scalable", "nonscalable", "allatonce", "allatonce_merged", "backend"}; 2262 PetscInt nalg = 5; 2263 #else 2264 const char *algTypes[6] = {"scalable", "nonscalable", "allatonce", "allatonce_merged", "backend", "hypre"}; 2265 PetscInt nalg = 6; 2266 #endif 2267 PetscInt pN = P->cmap->N; 2268 2269 PetscFunctionBegin; 2270 /* Check matrix local sizes */ 2271 PetscCall(PetscObjectGetComm((PetscObject)C, &comm)); 2272 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 ")", 2273 A->rmap->rstart, A->rmap->rend, P->rmap->rstart, P->rmap->rend); 2274 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 ")", 2275 A->cmap->rstart, A->cmap->rend, P->rmap->rstart, P->rmap->rend); 2276 2277 /* Set "nonscalable" as default algorithm */ 2278 PetscCall(PetscStrcmp(C->product->alg, "default", &flg)); 2279 if (flg) { 2280 PetscCall(MatProductSetAlgorithm(C, algTypes[alg])); 2281 2282 /* Set "scalable" as default if BN and local nonzeros of A and B are large */ 2283 if (pN > 100000) { 2284 MatInfo Ainfo, Pinfo; 2285 PetscInt nz_local; 2286 PetscBool alg_scalable_loc = PETSC_FALSE, alg_scalable; 2287 2288 PetscCall(MatGetInfo(A, MAT_LOCAL, &Ainfo)); 2289 PetscCall(MatGetInfo(P, MAT_LOCAL, &Pinfo)); 2290 nz_local = (PetscInt)(Ainfo.nz_allocated + Pinfo.nz_allocated); 2291 2292 if (pN > product->fill * nz_local) alg_scalable_loc = PETSC_TRUE; 2293 PetscCallMPI(MPIU_Allreduce(&alg_scalable_loc, &alg_scalable, 1, MPI_C_BOOL, MPI_LOR, comm)); 2294 2295 if (alg_scalable) { 2296 alg = 0; /* scalable algorithm would 50% slower than nonscalable algorithm */ 2297 PetscCall(MatProductSetAlgorithm(C, algTypes[alg])); 2298 } 2299 } 2300 } 2301 2302 /* Get runtime option */ 2303 if (product->api_user) { 2304 PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatPtAP", "Mat"); 2305 PetscCall(PetscOptionsEList("-matptap_via", "Algorithmic approach", "MatPtAP", algTypes, nalg, algTypes[alg], &alg, &flg)); 2306 PetscOptionsEnd(); 2307 } else { 2308 PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatProduct_PtAP", "Mat"); 2309 PetscCall(PetscOptionsEList("-mat_product_algorithm", "Algorithmic approach", "MatPtAP", algTypes, nalg, algTypes[alg], &alg, &flg)); 2310 PetscOptionsEnd(); 2311 } 2312 if (flg) PetscCall(MatProductSetAlgorithm(C, algTypes[alg])); 2313 2314 C->ops->productsymbolic = MatProductSymbolic_PtAP_MPIAIJ_MPIAIJ; 2315 PetscFunctionReturn(PETSC_SUCCESS); 2316 } 2317 2318 static PetscErrorCode MatProductSetFromOptions_MPIAIJ_RARt(Mat C) 2319 { 2320 Mat_Product *product = C->product; 2321 Mat A = product->A, R = product->B; 2322 2323 PetscFunctionBegin; 2324 /* Check matrix local sizes */ 2325 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, 2326 A->rmap->n, R->rmap->n, R->cmap->n); 2327 2328 C->ops->productsymbolic = MatProductSymbolic_RARt_MPIAIJ_MPIAIJ; 2329 PetscFunctionReturn(PETSC_SUCCESS); 2330 } 2331 2332 /* 2333 Set options for ABC = A*B*C = A*(B*C); ABC's algorithm must be chosen from AB's algorithm 2334 */ 2335 static PetscErrorCode MatProductSetFromOptions_MPIAIJ_ABC(Mat C) 2336 { 2337 Mat_Product *product = C->product; 2338 PetscBool flg = PETSC_FALSE; 2339 PetscInt alg = 1; /* default algorithm */ 2340 const char *algTypes[3] = {"scalable", "nonscalable", "seqmpi"}; 2341 PetscInt nalg = 3; 2342 2343 PetscFunctionBegin; 2344 /* Set default algorithm */ 2345 PetscCall(PetscStrcmp(C->product->alg, "default", &flg)); 2346 if (flg) PetscCall(MatProductSetAlgorithm(C, algTypes[alg])); 2347 2348 /* Get runtime option */ 2349 if (product->api_user) { 2350 PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatMatMatMult", "Mat"); 2351 PetscCall(PetscOptionsEList("-matmatmatmult_via", "Algorithmic approach", "MatMatMatMult", algTypes, nalg, algTypes[alg], &alg, &flg)); 2352 PetscOptionsEnd(); 2353 } else { 2354 PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatProduct_ABC", "Mat"); 2355 PetscCall(PetscOptionsEList("-mat_product_algorithm", "Algorithmic approach", "MatProduct_ABC", algTypes, nalg, algTypes[alg], &alg, &flg)); 2356 PetscOptionsEnd(); 2357 } 2358 if (flg) PetscCall(MatProductSetAlgorithm(C, algTypes[alg])); 2359 2360 C->ops->matmatmultsymbolic = MatMatMatMultSymbolic_MPIAIJ_MPIAIJ_MPIAIJ; 2361 C->ops->productsymbolic = MatProductSymbolic_ABC; 2362 PetscFunctionReturn(PETSC_SUCCESS); 2363 } 2364 2365 PETSC_INTERN PetscErrorCode MatProductSetFromOptions_MPIAIJ(Mat C) 2366 { 2367 Mat_Product *product = C->product; 2368 2369 PetscFunctionBegin; 2370 switch (product->type) { 2371 case MATPRODUCT_AB: 2372 PetscCall(MatProductSetFromOptions_MPIAIJ_AB(C)); 2373 break; 2374 case MATPRODUCT_ABt: 2375 PetscCall(MatProductSetFromOptions_MPIAIJ_ABt(C)); 2376 break; 2377 case MATPRODUCT_AtB: 2378 PetscCall(MatProductSetFromOptions_MPIAIJ_AtB(C)); 2379 break; 2380 case MATPRODUCT_PtAP: 2381 PetscCall(MatProductSetFromOptions_MPIAIJ_PtAP(C)); 2382 break; 2383 case MATPRODUCT_RARt: 2384 PetscCall(MatProductSetFromOptions_MPIAIJ_RARt(C)); 2385 break; 2386 case MATPRODUCT_ABC: 2387 PetscCall(MatProductSetFromOptions_MPIAIJ_ABC(C)); 2388 break; 2389 default: 2390 break; 2391 } 2392 PetscFunctionReturn(PETSC_SUCCESS); 2393 } 2394