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 MatProductCtxDestroy_MPIAIJ_MatMatMult(PetscCtxRt data) 77 { 78 MatProductCtx_APMPI *ptap = *(MatProductCtx_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 MatProductCtx_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 = (MatProductCtx_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 MatProductCtx_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 MatProductCtx_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(MatSetOption(C, MAT_NO_OFF_PROC_ENTRIES, PETSC_TRUE)); 312 PetscCall(MatAssemblyBegin(C, MAT_FINAL_ASSEMBLY)); 313 PetscCall(MatAssemblyEnd(C, MAT_FINAL_ASSEMBLY)); 314 PetscCall(MatSetOption(C, MAT_NEW_NONZERO_LOCATION_ERR, PETSC_TRUE)); 315 316 C->ops->matmultnumeric = MatMatMultNumeric_MPIAIJ_MPIAIJ_nonscalable; 317 C->ops->productnumeric = MatProductNumeric_AB; 318 319 /* attach the supporting struct to C for reuse */ 320 C->product->data = ptap; 321 C->product->destroy = MatProductCtxDestroy_MPIAIJ_MatMatMult; 322 323 /* set MatInfo */ 324 afill = (PetscReal)api[am] / (adi[am] + aoi[am] + pi_loc[pm] + 1) + 1.e-5; 325 if (afill < 1.0) afill = 1.0; 326 C->info.mallocs = nspacedouble; 327 C->info.fill_ratio_given = fill; 328 C->info.fill_ratio_needed = afill; 329 330 #if defined(PETSC_USE_INFO) 331 if (api[am]) { 332 PetscCall(PetscInfo(C, "Reallocs %" PetscInt_FMT "; Fill ratio: given %g needed %g.\n", nspacedouble, (double)fill, (double)afill)); 333 PetscCall(PetscInfo(C, "Use MatMatMult(A,B,MatReuse,%g,&C) for best performance.;\n", (double)afill)); 334 } else { 335 PetscCall(PetscInfo(C, "Empty matrix product\n")); 336 } 337 #endif 338 PetscFunctionReturn(PETSC_SUCCESS); 339 } 340 341 static PetscErrorCode MatMatMultSymbolic_MPIAIJ_MPIDense(Mat, Mat, PetscReal, Mat); 342 static PetscErrorCode MatMatMultNumeric_MPIAIJ_MPIDense(Mat, Mat, Mat); 343 344 static PetscErrorCode MatProductSetFromOptions_MPIAIJ_MPIDense_AB(Mat C) 345 { 346 Mat_Product *product = C->product; 347 Mat A = product->A, B = product->B; 348 349 PetscFunctionBegin; 350 if (A->cmap->rstart != B->rmap->rstart || A->cmap->rend != B->rmap->rend) 351 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); 352 353 C->ops->matmultsymbolic = MatMatMultSymbolic_MPIAIJ_MPIDense; 354 C->ops->productsymbolic = MatProductSymbolic_AB; 355 PetscFunctionReturn(PETSC_SUCCESS); 356 } 357 358 static PetscErrorCode MatProductSetFromOptions_MPIAIJ_MPIDense_AtB(Mat C) 359 { 360 Mat_Product *product = C->product; 361 Mat A = product->A, B = product->B; 362 363 PetscFunctionBegin; 364 if (A->rmap->rstart != B->rmap->rstart || A->rmap->rend != B->rmap->rend) 365 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); 366 367 C->ops->transposematmultsymbolic = MatTransposeMatMultSymbolic_MPIAIJ_MPIDense; 368 C->ops->productsymbolic = MatProductSymbolic_AtB; 369 PetscFunctionReturn(PETSC_SUCCESS); 370 } 371 372 PETSC_INTERN PetscErrorCode MatProductSetFromOptions_MPIAIJ_MPIDense(Mat C) 373 { 374 Mat_Product *product = C->product; 375 376 PetscFunctionBegin; 377 switch (product->type) { 378 case MATPRODUCT_AB: 379 PetscCall(MatProductSetFromOptions_MPIAIJ_MPIDense_AB(C)); 380 break; 381 case MATPRODUCT_AtB: 382 PetscCall(MatProductSetFromOptions_MPIAIJ_MPIDense_AtB(C)); 383 break; 384 default: 385 break; 386 } 387 PetscFunctionReturn(PETSC_SUCCESS); 388 } 389 390 typedef struct { 391 Mat workB, workB1; 392 MPI_Request *rwaits, *swaits; 393 PetscInt nsends, nrecvs; 394 MPI_Datatype *stype, *rtype; 395 PetscInt blda; 396 } MPIAIJ_MPIDense; 397 398 static PetscErrorCode MatMPIAIJ_MPIDenseDestroy(PetscCtxRt ctx) 399 { 400 MPIAIJ_MPIDense *contents = *(MPIAIJ_MPIDense **)ctx; 401 PetscInt i; 402 403 PetscFunctionBegin; 404 PetscCall(MatDestroy(&contents->workB)); 405 PetscCall(MatDestroy(&contents->workB1)); 406 for (i = 0; i < contents->nsends; i++) PetscCallMPI(MPI_Type_free(&contents->stype[i])); 407 for (i = 0; i < contents->nrecvs; i++) PetscCallMPI(MPI_Type_free(&contents->rtype[i])); 408 PetscCall(PetscFree4(contents->stype, contents->rtype, contents->rwaits, contents->swaits)); 409 PetscCall(PetscFree(contents)); 410 PetscFunctionReturn(PETSC_SUCCESS); 411 } 412 413 static PetscErrorCode MatMatMultSymbolic_MPIAIJ_MPIDense(Mat A, Mat B, PetscReal fill, Mat C) 414 { 415 Mat_MPIAIJ *aij = (Mat_MPIAIJ *)A->data; 416 PetscInt nz = aij->B->cmap->n, blda, m, M, n, N; 417 MPIAIJ_MPIDense *contents; 418 VecScatter ctx = aij->Mvctx; 419 PetscInt Am = A->rmap->n, Bm = B->rmap->n, BN = B->cmap->N, Bbn, Bbn1, bs, numBb; 420 MPI_Comm comm; 421 MPI_Datatype type1, *stype, *rtype; 422 const PetscInt *sindices, *sstarts, *rstarts; 423 PetscMPIInt *disp, nsends, nrecvs, nrows_to, nrows_from; 424 PetscBool cisdense; 425 426 PetscFunctionBegin; 427 MatCheckProduct(C, 4); 428 PetscCheck(!C->product->data, PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Product data not empty"); 429 PetscCall(PetscObjectGetComm((PetscObject)A, &comm)); 430 PetscCall(PetscObjectBaseTypeCompare((PetscObject)C, MATMPIDENSE, &cisdense)); 431 if (!cisdense) PetscCall(MatSetType(C, ((PetscObject)B)->type_name)); 432 PetscCall(MatGetLocalSize(C, &m, &n)); 433 PetscCall(MatGetSize(C, &M, &N)); 434 if (m == PETSC_DECIDE || n == PETSC_DECIDE || M == PETSC_DECIDE || N == PETSC_DECIDE) PetscCall(MatSetSizes(C, Am, B->cmap->n, A->rmap->N, BN)); 435 PetscCall(MatSetBlockSizesFromMats(C, A, B)); 436 PetscCall(MatSetUp(C)); 437 PetscCall(MatDenseGetLDA(B, &blda)); 438 PetscCall(PetscNew(&contents)); 439 440 PetscCall(VecScatterGetRemote_Private(ctx, PETSC_TRUE /*send*/, &nsends, &sstarts, &sindices, NULL, NULL)); 441 PetscCall(VecScatterGetRemoteOrdered_Private(ctx, PETSC_FALSE /*recv*/, &nrecvs, &rstarts, NULL, NULL, NULL)); 442 443 /* Create column block of B and C for memory scalability when BN is too large */ 444 /* Estimate Bbn, column size of Bb */ 445 if (nz) { 446 Bbn1 = 2 * Am * BN / nz; 447 if (!Bbn1) Bbn1 = 1; 448 } else Bbn1 = BN; 449 450 bs = B->cmap->bs; 451 Bbn1 = Bbn1 / bs * bs; /* Bbn1 is a multiple of bs */ 452 if (Bbn1 > BN) Bbn1 = BN; 453 PetscCallMPI(MPIU_Allreduce(&Bbn1, &Bbn, 1, MPIU_INT, MPI_MAX, comm)); 454 455 /* Enable runtime option for Bbn */ 456 PetscOptionsBegin(comm, ((PetscObject)C)->prefix, "MatProduct", "Mat"); 457 PetscCall(PetscOptionsDeprecated("-matmatmult_Bbn", "-matproduct_batch_size", "3.25", NULL)); 458 PetscCall(PetscOptionsInt("-matproduct_batch_size", "Number of columns in Bb", "MatProduct", Bbn, &Bbn, NULL)); 459 PetscOptionsEnd(); 460 Bbn = PetscMin(Bbn, BN); 461 462 if (Bbn > 0 && Bbn < BN) { 463 numBb = BN / Bbn; 464 Bbn1 = BN - numBb * Bbn; 465 } else numBb = 0; 466 467 if (numBb) { 468 PetscCall(PetscInfo(C, "use Bb, BN=%" PetscInt_FMT ", Bbn=%" PetscInt_FMT "; numBb=%" PetscInt_FMT "\n", BN, Bbn, numBb)); 469 if (Bbn1) { /* Create workB1 for the remaining columns */ 470 PetscCall(PetscInfo(C, "use Bb1, BN=%" PetscInt_FMT ", Bbn1=%" PetscInt_FMT "\n", BN, Bbn1)); 471 /* Create work matrix used to store off processor rows of B needed for local product */ 472 PetscCall(MatCreateSeqDense(PETSC_COMM_SELF, nz, Bbn1, NULL, &contents->workB1)); 473 } else contents->workB1 = NULL; 474 } 475 476 /* Create work matrix used to store off processor rows of B needed for local product */ 477 PetscCall(MatCreateSeqDense(PETSC_COMM_SELF, nz, Bbn, NULL, &contents->workB)); 478 479 /* Use MPI derived data type to reduce memory required by the send/recv buffers */ 480 PetscCall(PetscMalloc4(nsends, &stype, nrecvs, &rtype, nrecvs, &contents->rwaits, nsends, &contents->swaits)); 481 contents->stype = stype; 482 contents->nsends = nsends; 483 484 contents->rtype = rtype; 485 contents->nrecvs = nrecvs; 486 contents->blda = blda; 487 488 PetscCall(PetscMalloc1(Bm + 1, &disp)); 489 for (PetscMPIInt i = 0; i < nsends; i++) { 490 PetscCall(PetscMPIIntCast(sstarts[i + 1] - sstarts[i], &nrows_to)); 491 for (PetscInt j = 0; j < nrows_to; j++) PetscCall(PetscMPIIntCast(sindices[sstarts[i] + j], &disp[j])); /* rowB to be sent */ 492 PetscCallMPI(MPI_Type_create_indexed_block(nrows_to, 1, disp, MPIU_SCALAR, &type1)); 493 PetscCallMPI(MPI_Type_create_resized(type1, 0, blda * sizeof(PetscScalar), &stype[i])); 494 PetscCallMPI(MPI_Type_commit(&stype[i])); 495 PetscCallMPI(MPI_Type_free(&type1)); 496 } 497 498 for (PetscMPIInt i = 0; i < nrecvs; i++) { 499 /* received values from a process form a (nrows_from x Bbn) row block in workB (column-wise) */ 500 PetscCall(PetscMPIIntCast(rstarts[i + 1] - rstarts[i], &nrows_from)); 501 disp[0] = 0; 502 PetscCallMPI(MPI_Type_create_indexed_block(1, nrows_from, disp, MPIU_SCALAR, &type1)); 503 PetscCallMPI(MPI_Type_create_resized(type1, 0, nz * sizeof(PetscScalar), &rtype[i])); 504 PetscCallMPI(MPI_Type_commit(&rtype[i])); 505 PetscCallMPI(MPI_Type_free(&type1)); 506 } 507 508 PetscCall(PetscFree(disp)); 509 PetscCall(VecScatterRestoreRemote_Private(ctx, PETSC_TRUE /*send*/, &nsends, &sstarts, &sindices, NULL, NULL)); 510 PetscCall(VecScatterRestoreRemoteOrdered_Private(ctx, PETSC_FALSE /*recv*/, &nrecvs, &rstarts, NULL, NULL, NULL)); 511 PetscCall(MatSetOption(C, MAT_NO_OFF_PROC_ENTRIES, PETSC_TRUE)); 512 PetscCall(MatAssemblyBegin(C, MAT_FINAL_ASSEMBLY)); 513 PetscCall(MatAssemblyEnd(C, MAT_FINAL_ASSEMBLY)); 514 515 C->product->data = contents; 516 C->product->destroy = MatMPIAIJ_MPIDenseDestroy; 517 C->ops->matmultnumeric = MatMatMultNumeric_MPIAIJ_MPIDense; 518 PetscFunctionReturn(PETSC_SUCCESS); 519 } 520 521 PETSC_INTERN PetscErrorCode MatMatMultNumericAdd_SeqAIJ_SeqDense(Mat, Mat, Mat, const PetscBool); 522 523 /* 524 Performs an efficient scatter on the rows of B needed by this process; this is 525 a modification of the VecScatterBegin_() routines. 526 527 Input: If Bbidx = 0, uses B = Bb, else B = Bb1, see MatMatMultSymbolic_MPIAIJ_MPIDense() 528 */ 529 530 static PetscErrorCode MatMPIDenseScatter(Mat A, Mat B, PetscInt Bbidx, Mat C, Mat *outworkB) 531 { 532 Mat_MPIAIJ *aij = (Mat_MPIAIJ *)A->data; 533 const PetscScalar *b; 534 PetscScalar *rvalues; 535 VecScatter ctx = aij->Mvctx; 536 const PetscInt *sindices, *sstarts, *rstarts; 537 const PetscMPIInt *sprocs, *rprocs; 538 PetscMPIInt nsends, nrecvs; 539 MPI_Request *swaits, *rwaits; 540 MPI_Comm comm; 541 PetscMPIInt tag = ((PetscObject)ctx)->tag, ncols, nrows, nsends_mpi, nrecvs_mpi; 542 MPIAIJ_MPIDense *contents; 543 Mat workB; 544 MPI_Datatype *stype, *rtype; 545 PetscInt blda; 546 547 PetscFunctionBegin; 548 MatCheckProduct(C, 4); 549 PetscCheck(C->product->data, PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Product data empty"); 550 PetscCall(PetscMPIIntCast(B->cmap->N, &ncols)); 551 PetscCall(PetscMPIIntCast(aij->B->cmap->n, &nrows)); 552 contents = (MPIAIJ_MPIDense *)C->product->data; 553 PetscCall(VecScatterGetRemote_Private(ctx, PETSC_TRUE /*send*/, &nsends, &sstarts, &sindices, &sprocs, NULL /*bs*/)); 554 PetscCall(VecScatterGetRemoteOrdered_Private(ctx, PETSC_FALSE /*recv*/, &nrecvs, &rstarts, NULL, &rprocs, NULL /*bs*/)); 555 PetscCall(PetscMPIIntCast(nsends, &nsends_mpi)); 556 PetscCall(PetscMPIIntCast(nrecvs, &nrecvs_mpi)); 557 if (Bbidx == 0) workB = *outworkB = contents->workB; 558 else workB = *outworkB = contents->workB1; 559 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); 560 swaits = contents->swaits; 561 rwaits = contents->rwaits; 562 563 PetscCall(MatDenseGetArrayRead(B, &b)); 564 PetscCall(MatDenseGetLDA(B, &blda)); 565 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); 566 PetscCall(MatDenseGetArray(workB, &rvalues)); 567 568 /* Post recv, use MPI derived data type to save memory */ 569 PetscCall(PetscObjectGetComm((PetscObject)C, &comm)); 570 rtype = contents->rtype; 571 for (PetscMPIInt i = 0; i < nrecvs; i++) PetscCallMPI(MPIU_Irecv(rvalues + (rstarts[i] - rstarts[0]), ncols, rtype[i], rprocs[i], tag, comm, rwaits + i)); 572 573 stype = contents->stype; 574 for (PetscMPIInt i = 0; i < nsends; i++) PetscCallMPI(MPIU_Isend(b, ncols, stype[i], sprocs[i], tag, comm, swaits + i)); 575 576 if (nrecvs) PetscCallMPI(MPI_Waitall(nrecvs_mpi, rwaits, MPI_STATUSES_IGNORE)); 577 if (nsends) PetscCallMPI(MPI_Waitall(nsends_mpi, swaits, MPI_STATUSES_IGNORE)); 578 579 PetscCall(VecScatterRestoreRemote_Private(ctx, PETSC_TRUE /*send*/, &nsends, &sstarts, &sindices, &sprocs, NULL)); 580 PetscCall(VecScatterRestoreRemoteOrdered_Private(ctx, PETSC_FALSE /*recv*/, &nrecvs, &rstarts, NULL, &rprocs, NULL)); 581 PetscCall(MatDenseRestoreArrayRead(B, &b)); 582 PetscCall(MatDenseRestoreArray(workB, &rvalues)); 583 PetscFunctionReturn(PETSC_SUCCESS); 584 } 585 586 static PetscErrorCode MatMatMultNumeric_MPIAIJ_MPIDense(Mat A, Mat B, Mat C) 587 { 588 Mat_MPIAIJ *aij = (Mat_MPIAIJ *)A->data; 589 Mat_MPIDense *bdense = (Mat_MPIDense *)B->data; 590 Mat_MPIDense *cdense = (Mat_MPIDense *)C->data; 591 Mat workB; 592 MPIAIJ_MPIDense *contents; 593 594 PetscFunctionBegin; 595 MatCheckProduct(C, 3); 596 PetscCheck(C->product->data, PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Product data empty"); 597 contents = (MPIAIJ_MPIDense *)C->product->data; 598 /* diagonal block of A times all local rows of B */ 599 /* TODO: this calls a symbolic multiplication every time, which could be avoided */ 600 PetscCall(MatMatMult(aij->A, bdense->A, MAT_REUSE_MATRIX, PETSC_CURRENT, &cdense->A)); 601 if (contents->workB->cmap->n == B->cmap->N) { 602 /* get off processor parts of B needed to complete C=A*B */ 603 PetscCall(MatMPIDenseScatter(A, B, 0, C, &workB)); 604 605 /* off-diagonal block of A times nonlocal rows of B */ 606 PetscCall(MatMatMultNumericAdd_SeqAIJ_SeqDense(aij->B, workB, cdense->A, PETSC_TRUE)); 607 } else { 608 Mat Bb, Cb; 609 PetscInt BN = B->cmap->N, n = contents->workB->cmap->n; 610 PetscBool ccpu; 611 612 PetscCheck(n > 0, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Column block size %" PetscInt_FMT " must be positive", n); 613 /* Prevent from unneeded copies back and forth from the GPU 614 when getting and restoring the submatrix 615 We need a proper GPU code for AIJ * dense in parallel */ 616 PetscCall(MatBoundToCPU(C, &ccpu)); 617 PetscCall(MatBindToCPU(C, PETSC_TRUE)); 618 for (PetscInt i = 0; i < BN; i += n) { 619 PetscCall(MatDenseGetSubMatrix(B, PETSC_DECIDE, PETSC_DECIDE, i, PetscMin(i + n, BN), &Bb)); 620 PetscCall(MatDenseGetSubMatrix(C, PETSC_DECIDE, PETSC_DECIDE, i, PetscMin(i + n, BN), &Cb)); 621 622 /* get off processor parts of B needed to complete C=A*B */ 623 PetscCall(MatMPIDenseScatter(A, Bb, (i + n) > BN, C, &workB)); 624 625 /* off-diagonal block of A times nonlocal rows of B */ 626 cdense = (Mat_MPIDense *)Cb->data; 627 PetscCall(MatMatMultNumericAdd_SeqAIJ_SeqDense(aij->B, workB, cdense->A, PETSC_TRUE)); 628 PetscCall(MatDenseRestoreSubMatrix(B, &Bb)); 629 PetscCall(MatDenseRestoreSubMatrix(C, &Cb)); 630 } 631 PetscCall(MatBindToCPU(C, ccpu)); 632 } 633 PetscFunctionReturn(PETSC_SUCCESS); 634 } 635 636 PetscErrorCode MatMatMultNumeric_MPIAIJ_MPIAIJ(Mat A, Mat P, Mat C) 637 { 638 Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data, *c = (Mat_MPIAIJ *)C->data; 639 Mat_SeqAIJ *ad = (Mat_SeqAIJ *)a->A->data, *ao = (Mat_SeqAIJ *)a->B->data; 640 Mat_SeqAIJ *cd = (Mat_SeqAIJ *)c->A->data, *co = (Mat_SeqAIJ *)c->B->data; 641 PetscInt *adi = ad->i, *adj, *aoi = ao->i, *aoj; 642 PetscScalar *ada, *aoa, *cda = cd->a, *coa = co->a; 643 Mat_SeqAIJ *p_loc, *p_oth; 644 PetscInt *pi_loc, *pj_loc, *pi_oth, *pj_oth, *pj; 645 PetscScalar *pa_loc, *pa_oth, *pa, valtmp, *ca; 646 PetscInt cm = C->rmap->n, anz, pnz; 647 MatProductCtx_APMPI *ptap; 648 PetscScalar *apa_sparse; 649 const PetscScalar *dummy; 650 PetscInt *api, *apj, *apJ, i, j, k, row; 651 PetscInt cstart = C->cmap->rstart; 652 PetscInt cdnz, conz, k0, k1, nextp; 653 MPI_Comm comm; 654 PetscMPIInt size; 655 656 PetscFunctionBegin; 657 MatCheckProduct(C, 3); 658 ptap = (MatProductCtx_APMPI *)C->product->data; 659 PetscCheck(ptap, PetscObjectComm((PetscObject)C), PETSC_ERR_ARG_WRONGSTATE, "PtAP cannot be computed. Missing data"); 660 PetscCall(PetscObjectGetComm((PetscObject)C, &comm)); 661 PetscCallMPI(MPI_Comm_size(comm, &size)); 662 PetscCheck(ptap->P_oth || size <= 1, PetscObjectComm((PetscObject)C), PETSC_ERR_ARG_WRONGSTATE, "AP cannot be reused. Do not call MatProductClear()"); 663 664 /* flag CPU mask for C */ 665 #if defined(PETSC_HAVE_DEVICE) 666 if (C->offloadmask != PETSC_OFFLOAD_UNALLOCATED) C->offloadmask = PETSC_OFFLOAD_CPU; 667 if (c->A->offloadmask != PETSC_OFFLOAD_UNALLOCATED) c->A->offloadmask = PETSC_OFFLOAD_CPU; 668 if (c->B->offloadmask != PETSC_OFFLOAD_UNALLOCATED) c->B->offloadmask = PETSC_OFFLOAD_CPU; 669 #endif 670 apa_sparse = ptap->apa; 671 672 /* 1) get P_oth = ptap->P_oth and P_loc = ptap->P_loc */ 673 /* update numerical values of P_oth and P_loc */ 674 PetscCall(MatGetBrowsOfAoCols_MPIAIJ(A, P, MAT_REUSE_MATRIX, &ptap->startsj_s, &ptap->startsj_r, &ptap->bufa, &ptap->P_oth)); 675 PetscCall(MatMPIAIJGetLocalMat(P, MAT_REUSE_MATRIX, &ptap->P_loc)); 676 677 /* 2) compute numeric C_loc = A_loc*P = Ad*P_loc + Ao*P_oth */ 678 /* get data from symbolic products */ 679 p_loc = (Mat_SeqAIJ *)ptap->P_loc->data; 680 pi_loc = p_loc->i; 681 pj_loc = p_loc->j; 682 pa_loc = p_loc->a; 683 if (size > 1) { 684 p_oth = (Mat_SeqAIJ *)ptap->P_oth->data; 685 pi_oth = p_oth->i; 686 pj_oth = p_oth->j; 687 pa_oth = p_oth->a; 688 } else { 689 p_oth = NULL; 690 pi_oth = NULL; 691 pj_oth = NULL; 692 pa_oth = NULL; 693 } 694 695 /* trigger copy to CPU */ 696 PetscCall(MatSeqAIJGetArrayRead(a->A, &dummy)); 697 PetscCall(MatSeqAIJRestoreArrayRead(a->A, &dummy)); 698 PetscCall(MatSeqAIJGetArrayRead(a->B, &dummy)); 699 PetscCall(MatSeqAIJRestoreArrayRead(a->B, &dummy)); 700 api = ptap->api; 701 apj = ptap->apj; 702 for (i = 0; i < cm; i++) { 703 apJ = apj + api[i]; 704 705 /* diagonal portion of A */ 706 anz = adi[i + 1] - adi[i]; 707 adj = ad->j + adi[i]; 708 ada = ad->a + adi[i]; 709 for (j = 0; j < anz; j++) { 710 row = adj[j]; 711 pnz = pi_loc[row + 1] - pi_loc[row]; 712 pj = pj_loc + pi_loc[row]; 713 pa = pa_loc + pi_loc[row]; 714 /* perform sparse axpy */ 715 valtmp = ada[j]; 716 nextp = 0; 717 for (k = 0; nextp < pnz; k++) { 718 if (apJ[k] == pj[nextp]) { /* column of AP == column of P */ 719 apa_sparse[k] += valtmp * pa[nextp++]; 720 } 721 } 722 PetscCall(PetscLogFlops(2.0 * pnz)); 723 } 724 725 /* off-diagonal portion of A */ 726 anz = aoi[i + 1] - aoi[i]; 727 aoj = PetscSafePointerPlusOffset(ao->j, aoi[i]); 728 aoa = PetscSafePointerPlusOffset(ao->a, aoi[i]); 729 for (j = 0; j < anz; j++) { 730 row = aoj[j]; 731 pnz = pi_oth[row + 1] - pi_oth[row]; 732 pj = pj_oth + pi_oth[row]; 733 pa = pa_oth + pi_oth[row]; 734 /* perform sparse axpy */ 735 valtmp = aoa[j]; 736 nextp = 0; 737 for (k = 0; nextp < pnz; k++) { 738 if (apJ[k] == pj[nextp]) { /* column of AP == column of P */ 739 apa_sparse[k] += valtmp * pa[nextp++]; 740 } 741 } 742 PetscCall(PetscLogFlops(2.0 * pnz)); 743 } 744 745 /* set values in C */ 746 cdnz = cd->i[i + 1] - cd->i[i]; 747 conz = co->i[i + 1] - co->i[i]; 748 749 /* 1st off-diagonal part of C */ 750 ca = PetscSafePointerPlusOffset(coa, co->i[i]); 751 k = 0; 752 for (k0 = 0; k0 < conz; k0++) { 753 if (apJ[k] >= cstart) break; 754 ca[k0] = apa_sparse[k]; 755 apa_sparse[k] = 0.0; 756 k++; 757 } 758 759 /* diagonal part of C */ 760 ca = cda + cd->i[i]; 761 for (k1 = 0; k1 < cdnz; k1++) { 762 ca[k1] = apa_sparse[k]; 763 apa_sparse[k] = 0.0; 764 k++; 765 } 766 767 /* 2nd off-diagonal part of C */ 768 ca = PetscSafePointerPlusOffset(coa, co->i[i]); 769 for (; k0 < conz; k0++) { 770 ca[k0] = apa_sparse[k]; 771 apa_sparse[k] = 0.0; 772 k++; 773 } 774 } 775 PetscCall(MatAssemblyBegin(C, MAT_FINAL_ASSEMBLY)); 776 PetscCall(MatAssemblyEnd(C, MAT_FINAL_ASSEMBLY)); 777 PetscFunctionReturn(PETSC_SUCCESS); 778 } 779 780 /* same as MatMatMultSymbolic_MPIAIJ_MPIAIJ_nonscalable(), except using LLCondensed to avoid O(BN) memory requirement */ 781 PetscErrorCode MatMatMultSymbolic_MPIAIJ_MPIAIJ(Mat A, Mat P, PetscReal fill, Mat C) 782 { 783 MPI_Comm comm; 784 PetscMPIInt size; 785 MatProductCtx_APMPI *ptap; 786 PetscFreeSpaceList free_space = NULL, current_space = NULL; 787 Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data; 788 Mat_SeqAIJ *ad = (Mat_SeqAIJ *)a->A->data, *ao = (Mat_SeqAIJ *)a->B->data, *p_loc, *p_oth; 789 PetscInt *pi_loc, *pj_loc, *pi_oth, *pj_oth, *dnz, *onz; 790 PetscInt *adi = ad->i, *adj = ad->j, *aoi = ao->i, *aoj = ao->j, rstart = A->rmap->rstart; 791 PetscInt i, pnz, row, *api, *apj, *Jptr, apnz, nspacedouble = 0, j, nzi, *lnk, apnz_max = 1; 792 PetscInt am = A->rmap->n, pn = P->cmap->n, pm = P->rmap->n, lsize = pn + 20; 793 PetscReal afill; 794 MatType mtype; 795 796 PetscFunctionBegin; 797 MatCheckProduct(C, 4); 798 PetscCheck(!C->product->data, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Extra product struct not empty"); 799 PetscCall(PetscObjectGetComm((PetscObject)A, &comm)); 800 PetscCallMPI(MPI_Comm_size(comm, &size)); 801 802 /* create struct MatProductCtx_APMPI and attached it to C later */ 803 PetscCall(PetscNew(&ptap)); 804 805 /* get P_oth by taking rows of P (= non-zero cols of local A) from other processors */ 806 PetscCall(MatGetBrowsOfAoCols_MPIAIJ(A, P, MAT_INITIAL_MATRIX, &ptap->startsj_s, &ptap->startsj_r, &ptap->bufa, &ptap->P_oth)); 807 808 /* get P_loc by taking all local rows of P */ 809 PetscCall(MatMPIAIJGetLocalMat(P, MAT_INITIAL_MATRIX, &ptap->P_loc)); 810 811 p_loc = (Mat_SeqAIJ *)ptap->P_loc->data; 812 pi_loc = p_loc->i; 813 pj_loc = p_loc->j; 814 if (size > 1) { 815 p_oth = (Mat_SeqAIJ *)ptap->P_oth->data; 816 pi_oth = p_oth->i; 817 pj_oth = p_oth->j; 818 } else { 819 p_oth = NULL; 820 pi_oth = NULL; 821 pj_oth = NULL; 822 } 823 824 /* first, compute symbolic AP = A_loc*P = A_diag*P_loc + A_off*P_oth */ 825 PetscCall(PetscMalloc1(am + 1, &api)); 826 ptap->api = api; 827 api[0] = 0; 828 829 PetscCall(PetscLLCondensedCreate_Scalable(lsize, &lnk)); 830 831 /* Initial FreeSpace size is fill*(nnz(A)+nnz(P)) */ 832 PetscCall(PetscFreeSpaceGet(PetscRealIntMultTruncate(fill, PetscIntSumTruncate(adi[am], PetscIntSumTruncate(aoi[am], pi_loc[pm]))), &free_space)); 833 current_space = free_space; 834 MatPreallocateBegin(comm, am, pn, dnz, onz); 835 for (i = 0; i < am; i++) { 836 /* diagonal portion of A */ 837 nzi = adi[i + 1] - adi[i]; 838 for (j = 0; j < nzi; j++) { 839 row = *adj++; 840 pnz = pi_loc[row + 1] - pi_loc[row]; 841 Jptr = pj_loc + pi_loc[row]; 842 /* Expand list if it is not long enough */ 843 if (pnz + apnz_max > lsize) { 844 lsize = pnz + apnz_max; 845 PetscCall(PetscLLCondensedExpand_Scalable(lsize, &lnk)); 846 } 847 /* add non-zero cols of P into the sorted linked list lnk */ 848 PetscCall(PetscLLCondensedAddSorted_Scalable(pnz, Jptr, lnk)); 849 apnz = *lnk; /* The first element in the list is the number of items in the list */ 850 api[i + 1] = api[i] + apnz; 851 if (apnz > apnz_max) apnz_max = apnz + 1; /* '1' for diagonal entry */ 852 } 853 /* off-diagonal portion of A */ 854 nzi = aoi[i + 1] - aoi[i]; 855 for (j = 0; j < nzi; j++) { 856 row = *aoj++; 857 pnz = pi_oth[row + 1] - pi_oth[row]; 858 Jptr = pj_oth + pi_oth[row]; 859 /* Expand list if it is not long enough */ 860 if (pnz + apnz_max > lsize) { 861 lsize = pnz + apnz_max; 862 PetscCall(PetscLLCondensedExpand_Scalable(lsize, &lnk)); 863 } 864 /* add non-zero cols of P into the sorted linked list lnk */ 865 PetscCall(PetscLLCondensedAddSorted_Scalable(pnz, Jptr, lnk)); 866 apnz = *lnk; /* The first element in the list is the number of items in the list */ 867 api[i + 1] = api[i] + apnz; 868 if (apnz > apnz_max) apnz_max = apnz + 1; /* '1' for diagonal entry */ 869 } 870 871 /* add missing diagonal entry */ 872 if (C->force_diagonals) { 873 j = i + rstart; /* column index */ 874 PetscCall(PetscLLCondensedAddSorted_Scalable(1, &j, lnk)); 875 } 876 877 apnz = *lnk; 878 api[i + 1] = api[i] + apnz; 879 if (apnz > apnz_max) apnz_max = apnz; 880 881 /* if free space is not available, double the total space in the list */ 882 if (current_space->local_remaining < apnz) { 883 PetscCall(PetscFreeSpaceGet(PetscIntSumTruncate(apnz, current_space->total_array_size), ¤t_space)); 884 nspacedouble++; 885 } 886 887 /* Copy data into free space, then initialize lnk */ 888 PetscCall(PetscLLCondensedClean_Scalable(apnz, current_space->array, lnk)); 889 PetscCall(MatPreallocateSet(i + rstart, apnz, current_space->array, dnz, onz)); 890 891 current_space->array += apnz; 892 current_space->local_used += apnz; 893 current_space->local_remaining -= apnz; 894 } 895 896 /* Allocate space for apj, initialize apj, and */ 897 /* destroy list of free space and other temporary array(s) */ 898 PetscCall(PetscMalloc1(api[am], &ptap->apj)); 899 apj = ptap->apj; 900 PetscCall(PetscFreeSpaceContiguous(&free_space, ptap->apj)); 901 PetscCall(PetscLLCondensedDestroy_Scalable(lnk)); 902 903 /* create and assemble symbolic parallel matrix C */ 904 PetscCall(MatSetSizes(C, am, pn, PETSC_DETERMINE, PETSC_DETERMINE)); 905 PetscCall(MatSetBlockSizesFromMats(C, A, P)); 906 PetscCall(MatGetType(A, &mtype)); 907 PetscCall(MatSetType(C, mtype)); 908 PetscCall(MatMPIAIJSetPreallocation(C, 0, dnz, 0, onz)); 909 MatPreallocateEnd(dnz, onz); 910 911 /* malloc apa for assembly C */ 912 PetscCall(PetscCalloc1(apnz_max, &ptap->apa)); 913 914 PetscCall(MatSetValues_MPIAIJ_CopyFromCSRFormat_Symbolic(C, apj, api)); 915 PetscCall(MatSetOption(C, MAT_NO_OFF_PROC_ENTRIES, PETSC_TRUE)); 916 PetscCall(MatAssemblyBegin(C, MAT_FINAL_ASSEMBLY)); 917 PetscCall(MatAssemblyEnd(C, MAT_FINAL_ASSEMBLY)); 918 PetscCall(MatSetOption(C, MAT_NEW_NONZERO_LOCATION_ERR, PETSC_TRUE)); 919 920 C->ops->matmultnumeric = MatMatMultNumeric_MPIAIJ_MPIAIJ; 921 C->ops->productnumeric = MatProductNumeric_AB; 922 923 /* attach the supporting struct to C for reuse */ 924 C->product->data = ptap; 925 C->product->destroy = MatProductCtxDestroy_MPIAIJ_MatMatMult; 926 927 /* set MatInfo */ 928 afill = (PetscReal)api[am] / (adi[am] + aoi[am] + pi_loc[pm] + 1) + 1.e-5; 929 if (afill < 1.0) afill = 1.0; 930 C->info.mallocs = nspacedouble; 931 C->info.fill_ratio_given = fill; 932 C->info.fill_ratio_needed = afill; 933 934 #if defined(PETSC_USE_INFO) 935 if (api[am]) { 936 PetscCall(PetscInfo(C, "Reallocs %" PetscInt_FMT "; Fill ratio: given %g needed %g.\n", nspacedouble, (double)fill, (double)afill)); 937 PetscCall(PetscInfo(C, "Use MatMatMult(A,B,MatReuse,%g,&C) for best performance.;\n", (double)afill)); 938 } else { 939 PetscCall(PetscInfo(C, "Empty matrix product\n")); 940 } 941 #endif 942 PetscFunctionReturn(PETSC_SUCCESS); 943 } 944 945 /* This function is needed for the seqMPI matrix-matrix multiplication. */ 946 /* Three input arrays are merged to one output array. The size of the */ 947 /* output array is also output. Duplicate entries only show up once. */ 948 static void Merge3SortedArrays(PetscInt size1, PetscInt *in1, PetscInt size2, PetscInt *in2, PetscInt size3, PetscInt *in3, PetscInt *size4, PetscInt *out) 949 { 950 int i = 0, j = 0, k = 0, l = 0; 951 952 /* Traverse all three arrays */ 953 while (i < size1 && j < size2 && k < size3) { 954 if (in1[i] < in2[j] && in1[i] < in3[k]) { 955 out[l++] = in1[i++]; 956 } else if (in2[j] < in1[i] && in2[j] < in3[k]) { 957 out[l++] = in2[j++]; 958 } else if (in3[k] < in1[i] && in3[k] < in2[j]) { 959 out[l++] = in3[k++]; 960 } else if (in1[i] == in2[j] && in1[i] < in3[k]) { 961 out[l++] = in1[i]; 962 i++, j++; 963 } else if (in1[i] == in3[k] && in1[i] < in2[j]) { 964 out[l++] = in1[i]; 965 i++, k++; 966 } else if (in3[k] == in2[j] && in2[j] < in1[i]) { 967 out[l++] = in2[j]; 968 k++, j++; 969 } else if (in1[i] == in2[j] && in1[i] == in3[k]) { 970 out[l++] = in1[i]; 971 i++, j++, k++; 972 } 973 } 974 975 /* Traverse two remaining arrays */ 976 while (i < size1 && j < size2) { 977 if (in1[i] < in2[j]) { 978 out[l++] = in1[i++]; 979 } else if (in1[i] > in2[j]) { 980 out[l++] = in2[j++]; 981 } else { 982 out[l++] = in1[i]; 983 i++, j++; 984 } 985 } 986 987 while (i < size1 && k < size3) { 988 if (in1[i] < in3[k]) { 989 out[l++] = in1[i++]; 990 } else if (in1[i] > in3[k]) { 991 out[l++] = in3[k++]; 992 } else { 993 out[l++] = in1[i]; 994 i++, k++; 995 } 996 } 997 998 while (k < size3 && j < size2) { 999 if (in3[k] < in2[j]) { 1000 out[l++] = in3[k++]; 1001 } else if (in3[k] > in2[j]) { 1002 out[l++] = in2[j++]; 1003 } else { 1004 out[l++] = in3[k]; 1005 k++, j++; 1006 } 1007 } 1008 1009 /* Traverse one remaining array */ 1010 while (i < size1) out[l++] = in1[i++]; 1011 while (j < size2) out[l++] = in2[j++]; 1012 while (k < size3) out[l++] = in3[k++]; 1013 1014 *size4 = l; 1015 } 1016 1017 /* This matrix-matrix multiplication algorithm divides the multiplication into three multiplications and */ 1018 /* adds up the products. Two of these three multiplications are performed with existing (sequential) */ 1019 /* matrix-matrix multiplications. */ 1020 PetscErrorCode MatMatMultSymbolic_MPIAIJ_MPIAIJ_seqMPI(Mat A, Mat P, PetscReal fill, Mat C) 1021 { 1022 MPI_Comm comm; 1023 PetscMPIInt size; 1024 MatProductCtx_APMPI *ptap; 1025 PetscFreeSpaceList free_space_diag = NULL, current_space = NULL; 1026 Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data; 1027 Mat_SeqAIJ *ad = (Mat_SeqAIJ *)a->A->data, *ao = (Mat_SeqAIJ *)a->B->data, *p_loc; 1028 Mat_MPIAIJ *p = (Mat_MPIAIJ *)P->data; 1029 Mat_SeqAIJ *adpd_seq, *p_off, *aopoth_seq; 1030 PetscInt adponz, adpdnz; 1031 PetscInt *pi_loc, *dnz, *onz; 1032 PetscInt *adi = ad->i, *adj = ad->j, *aoi = ao->i, rstart = A->rmap->rstart; 1033 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; 1034 PetscInt am = A->rmap->n, pN = P->cmap->N, pn = P->cmap->n, pm = P->rmap->n, p_colstart, p_colend; 1035 PetscBT lnkbt; 1036 PetscReal afill; 1037 PetscMPIInt rank; 1038 Mat adpd, aopoth; 1039 MatType mtype; 1040 const char *prefix; 1041 1042 PetscFunctionBegin; 1043 MatCheckProduct(C, 4); 1044 PetscCheck(!C->product->data, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Extra product struct not empty"); 1045 PetscCall(PetscObjectGetComm((PetscObject)A, &comm)); 1046 PetscCallMPI(MPI_Comm_size(comm, &size)); 1047 PetscCallMPI(MPI_Comm_rank(comm, &rank)); 1048 PetscCall(MatGetOwnershipRangeColumn(P, &p_colstart, &p_colend)); 1049 1050 /* create struct MatProductCtx_APMPI and attached it to C later */ 1051 PetscCall(PetscNew(&ptap)); 1052 1053 /* get P_oth by taking rows of P (= non-zero cols of local A) from other processors */ 1054 PetscCall(MatGetBrowsOfAoCols_MPIAIJ(A, P, MAT_INITIAL_MATRIX, &ptap->startsj_s, &ptap->startsj_r, &ptap->bufa, &ptap->P_oth)); 1055 1056 /* get P_loc by taking all local rows of P */ 1057 PetscCall(MatMPIAIJGetLocalMat(P, MAT_INITIAL_MATRIX, &ptap->P_loc)); 1058 1059 p_loc = (Mat_SeqAIJ *)ptap->P_loc->data; 1060 pi_loc = p_loc->i; 1061 1062 /* Allocate memory for the i arrays of the matrices A*P, A_diag*P_off and A_offd * P */ 1063 PetscCall(PetscMalloc1(am + 1, &api)); 1064 PetscCall(PetscMalloc1(am + 1, &adpoi)); 1065 1066 adpoi[0] = 0; 1067 ptap->api = api; 1068 api[0] = 0; 1069 1070 /* create and initialize a linked list, will be used for both A_diag * P_loc_off and A_offd * P_oth */ 1071 PetscCall(PetscLLCondensedCreate(pN, pN, &lnk, &lnkbt)); 1072 MatPreallocateBegin(comm, am, pn, dnz, onz); 1073 1074 /* Symbolic calc of A_loc_diag * P_loc_diag */ 1075 PetscCall(MatGetOptionsPrefix(A, &prefix)); 1076 PetscCall(MatProductCreate(a->A, p->A, NULL, &adpd)); 1077 PetscCall(MatGetOptionsPrefix(A, &prefix)); 1078 PetscCall(MatSetOptionsPrefix(adpd, prefix)); 1079 PetscCall(MatAppendOptionsPrefix(adpd, "inner_diag_")); 1080 1081 PetscCall(MatProductSetType(adpd, MATPRODUCT_AB)); 1082 PetscCall(MatProductSetAlgorithm(adpd, "sorted")); 1083 PetscCall(MatProductSetFill(adpd, fill)); 1084 PetscCall(MatProductSetFromOptions(adpd)); 1085 1086 adpd->force_diagonals = C->force_diagonals; 1087 PetscCall(MatProductSymbolic(adpd)); 1088 1089 adpd_seq = (Mat_SeqAIJ *)((adpd)->data); 1090 adpdi = adpd_seq->i; 1091 adpdj = adpd_seq->j; 1092 p_off = (Mat_SeqAIJ *)p->B->data; 1093 poff_i = p_off->i; 1094 poff_j = p_off->j; 1095 1096 /* j_temp stores indices of a result row before they are added to the linked list */ 1097 PetscCall(PetscMalloc1(pN, &j_temp)); 1098 1099 /* Symbolic calc of the A_diag * p_loc_off */ 1100 /* Initial FreeSpace size is fill*(nnz(A)+nnz(P)) */ 1101 PetscCall(PetscFreeSpaceGet(PetscRealIntMultTruncate(fill, PetscIntSumTruncate(adi[am], PetscIntSumTruncate(aoi[am], pi_loc[pm]))), &free_space_diag)); 1102 current_space = free_space_diag; 1103 1104 for (i = 0; i < am; i++) { 1105 /* A_diag * P_loc_off */ 1106 nzi = adi[i + 1] - adi[i]; 1107 for (j = 0; j < nzi; j++) { 1108 row = *adj++; 1109 pnz = poff_i[row + 1] - poff_i[row]; 1110 Jptr = poff_j + poff_i[row]; 1111 for (i1 = 0; i1 < pnz; i1++) j_temp[i1] = p->garray[Jptr[i1]]; 1112 /* add non-zero cols of P into the sorted linked list lnk */ 1113 PetscCall(PetscLLCondensedAddSorted(pnz, j_temp, lnk, lnkbt)); 1114 } 1115 1116 adponz = lnk[0]; 1117 adpoi[i + 1] = adpoi[i] + adponz; 1118 1119 /* if free space is not available, double the total space in the list */ 1120 if (current_space->local_remaining < adponz) { 1121 PetscCall(PetscFreeSpaceGet(PetscIntSumTruncate(adponz, current_space->total_array_size), ¤t_space)); 1122 nspacedouble++; 1123 } 1124 1125 /* Copy data into free space, then initialize lnk */ 1126 PetscCall(PetscLLCondensedClean(pN, adponz, current_space->array, lnk, lnkbt)); 1127 1128 current_space->array += adponz; 1129 current_space->local_used += adponz; 1130 current_space->local_remaining -= adponz; 1131 } 1132 1133 /* Symbolic calc of A_off * P_oth */ 1134 PetscCall(MatSetOptionsPrefix(a->B, prefix)); 1135 PetscCall(MatAppendOptionsPrefix(a->B, "inner_offdiag_")); 1136 PetscCall(MatCreate(PETSC_COMM_SELF, &aopoth)); 1137 PetscCall(MatMatMultSymbolic_SeqAIJ_SeqAIJ(a->B, ptap->P_oth, fill, aopoth)); 1138 aopoth_seq = (Mat_SeqAIJ *)((aopoth)->data); 1139 aopothi = aopoth_seq->i; 1140 aopothj = aopoth_seq->j; 1141 1142 /* Allocate space for apj, adpj, aopj, ... */ 1143 /* destroy lists of free space and other temporary array(s) */ 1144 1145 PetscCall(PetscMalloc1(aopothi[am] + adpoi[am] + adpdi[am], &ptap->apj)); 1146 PetscCall(PetscMalloc1(adpoi[am], &adpoj)); 1147 1148 /* Copy from linked list to j-array */ 1149 PetscCall(PetscFreeSpaceContiguous(&free_space_diag, adpoj)); 1150 PetscCall(PetscLLDestroy(lnk, lnkbt)); 1151 1152 adpoJ = adpoj; 1153 adpdJ = adpdj; 1154 aopJ = aopothj; 1155 apj = ptap->apj; 1156 apJ = apj; /* still empty */ 1157 1158 /* Merge j-arrays of A_off * P, A_diag * P_loc_off, and */ 1159 /* A_diag * P_loc_diag to get A*P */ 1160 for (i = 0; i < am; i++) { 1161 aopnz = aopothi[i + 1] - aopothi[i]; 1162 adponz = adpoi[i + 1] - adpoi[i]; 1163 adpdnz = adpdi[i + 1] - adpdi[i]; 1164 1165 /* Correct indices from A_diag*P_diag */ 1166 for (i1 = 0; i1 < adpdnz; i1++) adpdJ[i1] += p_colstart; 1167 /* Merge j-arrays of A_diag * P_loc_off and A_diag * P_loc_diag and A_off * P_oth */ 1168 Merge3SortedArrays(adponz, adpoJ, adpdnz, adpdJ, aopnz, aopJ, &apnz, apJ); 1169 PetscCall(MatPreallocateSet(i + rstart, apnz, apJ, dnz, onz)); 1170 1171 aopJ += aopnz; 1172 adpoJ += adponz; 1173 adpdJ += adpdnz; 1174 apJ += apnz; 1175 api[i + 1] = api[i] + apnz; 1176 } 1177 1178 /* malloc apa to store dense row A[i,:]*P */ 1179 PetscCall(PetscCalloc1(pN, &ptap->apa)); 1180 1181 /* create and assemble symbolic parallel matrix C */ 1182 PetscCall(MatSetSizes(C, am, pn, PETSC_DETERMINE, PETSC_DETERMINE)); 1183 PetscCall(MatSetBlockSizesFromMats(C, A, P)); 1184 PetscCall(MatGetType(A, &mtype)); 1185 PetscCall(MatSetType(C, mtype)); 1186 PetscCall(MatMPIAIJSetPreallocation(C, 0, dnz, 0, onz)); 1187 MatPreallocateEnd(dnz, onz); 1188 1189 PetscCall(MatSetValues_MPIAIJ_CopyFromCSRFormat_Symbolic(C, apj, api)); 1190 PetscCall(MatSetOption(C, MAT_NO_OFF_PROC_ENTRIES, PETSC_TRUE)); 1191 PetscCall(MatAssemblyBegin(C, MAT_FINAL_ASSEMBLY)); 1192 PetscCall(MatAssemblyEnd(C, MAT_FINAL_ASSEMBLY)); 1193 PetscCall(MatSetOption(C, MAT_NEW_NONZERO_LOCATION_ERR, PETSC_TRUE)); 1194 1195 C->ops->matmultnumeric = MatMatMultNumeric_MPIAIJ_MPIAIJ_nonscalable; 1196 C->ops->productnumeric = MatProductNumeric_AB; 1197 1198 /* attach the supporting struct to C for reuse */ 1199 C->product->data = ptap; 1200 C->product->destroy = MatProductCtxDestroy_MPIAIJ_MatMatMult; 1201 1202 /* set MatInfo */ 1203 afill = (PetscReal)api[am] / (adi[am] + aoi[am] + pi_loc[pm] + 1) + 1.e-5; 1204 if (afill < 1.0) afill = 1.0; 1205 C->info.mallocs = nspacedouble; 1206 C->info.fill_ratio_given = fill; 1207 C->info.fill_ratio_needed = afill; 1208 1209 #if defined(PETSC_USE_INFO) 1210 if (api[am]) { 1211 PetscCall(PetscInfo(C, "Reallocs %" PetscInt_FMT "; Fill ratio: given %g needed %g.\n", nspacedouble, (double)fill, (double)afill)); 1212 PetscCall(PetscInfo(C, "Use MatMatMult(A,B,MatReuse,%g,&C) for best performance.;\n", (double)afill)); 1213 } else { 1214 PetscCall(PetscInfo(C, "Empty matrix product\n")); 1215 } 1216 #endif 1217 1218 PetscCall(MatDestroy(&aopoth)); 1219 PetscCall(MatDestroy(&adpd)); 1220 PetscCall(PetscFree(j_temp)); 1221 PetscCall(PetscFree(adpoj)); 1222 PetscCall(PetscFree(adpoi)); 1223 PetscFunctionReturn(PETSC_SUCCESS); 1224 } 1225 1226 /* This routine only works when scall=MAT_REUSE_MATRIX! */ 1227 PetscErrorCode MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ_matmatmult(Mat P, Mat A, Mat C) 1228 { 1229 MatProductCtx_APMPI *ptap; 1230 Mat Pt; 1231 1232 PetscFunctionBegin; 1233 MatCheckProduct(C, 3); 1234 ptap = (MatProductCtx_APMPI *)C->product->data; 1235 PetscCheck(ptap, PetscObjectComm((PetscObject)C), PETSC_ERR_ARG_WRONGSTATE, "PtAP cannot be computed. Missing data"); 1236 PetscCheck(ptap->Pt, PetscObjectComm((PetscObject)C), PETSC_ERR_ARG_WRONGSTATE, "PtA cannot be reused. Do not call MatProductClear()"); 1237 1238 Pt = ptap->Pt; 1239 PetscCall(MatTransposeSetPrecursor(P, Pt)); 1240 PetscCall(MatTranspose(P, MAT_REUSE_MATRIX, &Pt)); 1241 PetscCall(MatMatMultNumeric_MPIAIJ_MPIAIJ(Pt, A, C)); 1242 PetscFunctionReturn(PETSC_SUCCESS); 1243 } 1244 1245 /* This routine is modified from MatPtAPSymbolic_MPIAIJ_MPIAIJ() */ 1246 PetscErrorCode MatTransposeMatMultSymbolic_MPIAIJ_MPIAIJ_nonscalable(Mat P, Mat A, PetscReal fill, Mat C) 1247 { 1248 MatProductCtx_APMPI *ptap; 1249 Mat_MPIAIJ *p = (Mat_MPIAIJ *)P->data; 1250 MPI_Comm comm; 1251 PetscMPIInt size, rank; 1252 PetscFreeSpaceList free_space = NULL, current_space = NULL; 1253 PetscInt pn = P->cmap->n, aN = A->cmap->N, an = A->cmap->n; 1254 PetscInt *lnk, i, k, rstart; 1255 PetscBT lnkbt; 1256 PetscMPIInt tagi, tagj, *len_si, *len_s, *len_ri, nrecv, proc, nsend; 1257 PETSC_UNUSED PetscMPIInt icompleted = 0; 1258 PetscInt **buf_rj, **buf_ri, **buf_ri_k, row, ncols, *cols; 1259 PetscInt len, *dnz, *onz, *owners, nzi; 1260 PetscInt nrows, *buf_s, *buf_si, *buf_si_i, **nextrow, **nextci; 1261 MPI_Request *swaits, *rwaits; 1262 MPI_Status *sstatus, rstatus; 1263 PetscLayout rowmap; 1264 PetscInt *owners_co, *coi, *coj; /* i and j array of (p->B)^T*A*P - used in the communication */ 1265 PetscMPIInt *len_r, *id_r; /* array of length of comm->size, store send/recv matrix values */ 1266 PetscInt *Jptr, *prmap = p->garray, con, j, Crmax; 1267 Mat_SeqAIJ *a_loc, *c_loc, *c_oth; 1268 PetscHMapI ta; 1269 MatType mtype; 1270 const char *prefix; 1271 1272 PetscFunctionBegin; 1273 PetscCall(PetscObjectGetComm((PetscObject)A, &comm)); 1274 PetscCallMPI(MPI_Comm_size(comm, &size)); 1275 PetscCallMPI(MPI_Comm_rank(comm, &rank)); 1276 1277 /* create symbolic parallel matrix C */ 1278 PetscCall(MatGetType(A, &mtype)); 1279 PetscCall(MatSetType(C, mtype)); 1280 1281 C->ops->transposematmultnumeric = MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ_nonscalable; 1282 1283 /* create struct MatProductCtx_APMPI and attached it to C later */ 1284 PetscCall(PetscNew(&ptap)); 1285 1286 /* (0) compute Rd = Pd^T, Ro = Po^T */ 1287 PetscCall(MatTranspose(p->A, MAT_INITIAL_MATRIX, &ptap->Rd)); 1288 PetscCall(MatTranspose(p->B, MAT_INITIAL_MATRIX, &ptap->Ro)); 1289 1290 /* (1) compute symbolic A_loc */ 1291 PetscCall(MatMPIAIJGetLocalMat(A, MAT_INITIAL_MATRIX, &ptap->A_loc)); 1292 1293 /* (2-1) compute symbolic C_oth = Ro*A_loc */ 1294 PetscCall(MatGetOptionsPrefix(A, &prefix)); 1295 PetscCall(MatSetOptionsPrefix(ptap->Ro, prefix)); 1296 PetscCall(MatAppendOptionsPrefix(ptap->Ro, "inner_offdiag_")); 1297 PetscCall(MatCreate(PETSC_COMM_SELF, &ptap->C_oth)); 1298 PetscCall(MatMatMultSymbolic_SeqAIJ_SeqAIJ(ptap->Ro, ptap->A_loc, fill, ptap->C_oth)); 1299 1300 /* (3) send coj of C_oth to other processors */ 1301 /* determine row ownership */ 1302 PetscCall(PetscLayoutCreate(comm, &rowmap)); 1303 rowmap->n = pn; 1304 rowmap->bs = 1; 1305 PetscCall(PetscLayoutSetUp(rowmap)); 1306 owners = rowmap->range; 1307 1308 /* determine the number of messages to send, their lengths */ 1309 PetscCall(PetscMalloc4(size, &len_s, size, &len_si, size, &sstatus, size + 1, &owners_co)); 1310 PetscCall(PetscArrayzero(len_s, size)); 1311 PetscCall(PetscArrayzero(len_si, size)); 1312 1313 c_oth = (Mat_SeqAIJ *)ptap->C_oth->data; 1314 coi = c_oth->i; 1315 coj = c_oth->j; 1316 con = ptap->C_oth->rmap->n; 1317 proc = 0; 1318 for (i = 0; i < con; i++) { 1319 while (prmap[i] >= owners[proc + 1]) proc++; 1320 len_si[proc]++; /* num of rows in Co(=Pt*A) to be sent to [proc] */ 1321 len_s[proc] += coi[i + 1] - coi[i]; /* num of nonzeros in Co to be sent to [proc] */ 1322 } 1323 1324 len = 0; /* max length of buf_si[], see (4) */ 1325 owners_co[0] = 0; 1326 nsend = 0; 1327 for (proc = 0; proc < size; proc++) { 1328 owners_co[proc + 1] = owners_co[proc] + len_si[proc]; 1329 if (len_s[proc]) { 1330 nsend++; 1331 len_si[proc] = 2 * (len_si[proc] + 1); /* length of buf_si to be sent to [proc] */ 1332 len += len_si[proc]; 1333 } 1334 } 1335 1336 /* determine the number and length of messages to receive for coi and coj */ 1337 PetscCall(PetscGatherNumberOfMessages(comm, NULL, len_s, &nrecv)); 1338 PetscCall(PetscGatherMessageLengths2(comm, nsend, nrecv, len_s, len_si, &id_r, &len_r, &len_ri)); 1339 1340 /* post the Irecv and Isend of coj */ 1341 PetscCall(PetscCommGetNewTag(comm, &tagj)); 1342 PetscCall(PetscPostIrecvInt(comm, tagj, nrecv, id_r, len_r, &buf_rj, &rwaits)); 1343 PetscCall(PetscMalloc1(nsend, &swaits)); 1344 for (proc = 0, k = 0; proc < size; proc++) { 1345 if (!len_s[proc]) continue; 1346 i = owners_co[proc]; 1347 PetscCallMPI(MPIU_Isend(coj + coi[i], len_s[proc], MPIU_INT, proc, tagj, comm, swaits + k)); 1348 k++; 1349 } 1350 1351 /* (2-2) compute symbolic C_loc = Rd*A_loc */ 1352 PetscCall(MatSetOptionsPrefix(ptap->Rd, prefix)); 1353 PetscCall(MatAppendOptionsPrefix(ptap->Rd, "inner_diag_")); 1354 PetscCall(MatCreate(PETSC_COMM_SELF, &ptap->C_loc)); 1355 PetscCall(MatMatMultSymbolic_SeqAIJ_SeqAIJ(ptap->Rd, ptap->A_loc, fill, ptap->C_loc)); 1356 c_loc = (Mat_SeqAIJ *)ptap->C_loc->data; 1357 1358 /* receives coj are complete */ 1359 for (i = 0; i < nrecv; i++) PetscCallMPI(MPI_Waitany(nrecv, rwaits, &icompleted, &rstatus)); 1360 PetscCall(PetscFree(rwaits)); 1361 if (nsend) PetscCallMPI(MPI_Waitall(nsend, swaits, sstatus)); 1362 1363 /* add received column indices into ta to update Crmax */ 1364 a_loc = (Mat_SeqAIJ *)ptap->A_loc->data; 1365 1366 /* create and initialize a linked list */ 1367 PetscCall(PetscHMapICreateWithSize(an, &ta)); /* for compute Crmax */ 1368 MatRowMergeMax_SeqAIJ(a_loc, ptap->A_loc->rmap->N, ta); 1369 1370 for (k = 0; k < nrecv; k++) { /* k-th received message */ 1371 Jptr = buf_rj[k]; 1372 for (j = 0; j < len_r[k]; j++) PetscCall(PetscHMapISet(ta, *(Jptr + j) + 1, 1)); 1373 } 1374 PetscCall(PetscHMapIGetSize(ta, &Crmax)); 1375 PetscCall(PetscHMapIDestroy(&ta)); 1376 1377 /* (4) send and recv coi */ 1378 PetscCall(PetscCommGetNewTag(comm, &tagi)); 1379 PetscCall(PetscPostIrecvInt(comm, tagi, nrecv, id_r, len_ri, &buf_ri, &rwaits)); 1380 PetscCall(PetscMalloc1(len, &buf_s)); 1381 buf_si = buf_s; /* points to the beginning of k-th msg to be sent */ 1382 for (proc = 0, k = 0; proc < size; proc++) { 1383 if (!len_s[proc]) continue; 1384 /* form outgoing message for i-structure: 1385 buf_si[0]: nrows to be sent 1386 [1:nrows]: row index (global) 1387 [nrows+1:2*nrows+1]: i-structure index 1388 */ 1389 nrows = len_si[proc] / 2 - 1; /* num of rows in Co to be sent to [proc] */ 1390 buf_si_i = buf_si + nrows + 1; 1391 buf_si[0] = nrows; 1392 buf_si_i[0] = 0; 1393 nrows = 0; 1394 for (i = owners_co[proc]; i < owners_co[proc + 1]; i++) { 1395 nzi = coi[i + 1] - coi[i]; 1396 buf_si_i[nrows + 1] = buf_si_i[nrows] + nzi; /* i-structure */ 1397 buf_si[nrows + 1] = prmap[i] - owners[proc]; /* local row index */ 1398 nrows++; 1399 } 1400 PetscCallMPI(MPIU_Isend(buf_si, len_si[proc], MPIU_INT, proc, tagi, comm, swaits + k)); 1401 k++; 1402 buf_si += len_si[proc]; 1403 } 1404 for (i = 0; i < nrecv; i++) PetscCallMPI(MPI_Waitany(nrecv, rwaits, &icompleted, &rstatus)); 1405 PetscCall(PetscFree(rwaits)); 1406 if (nsend) PetscCallMPI(MPI_Waitall(nsend, swaits, sstatus)); 1407 1408 PetscCall(PetscFree4(len_s, len_si, sstatus, owners_co)); 1409 PetscCall(PetscFree(len_ri)); 1410 PetscCall(PetscFree(swaits)); 1411 PetscCall(PetscFree(buf_s)); 1412 1413 /* (5) compute the local portion of C */ 1414 /* set initial free space to be Crmax, sufficient for holding nonzeros in each row of C */ 1415 PetscCall(PetscFreeSpaceGet(Crmax, &free_space)); 1416 current_space = free_space; 1417 1418 PetscCall(PetscMalloc3(nrecv, &buf_ri_k, nrecv, &nextrow, nrecv, &nextci)); 1419 for (k = 0; k < nrecv; k++) { 1420 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 1421 nrows = *buf_ri_k[k]; 1422 nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */ 1423 nextci[k] = buf_ri_k[k] + (nrows + 1); /* points to the next i-structure of k-th recved i-structure */ 1424 } 1425 1426 MatPreallocateBegin(comm, pn, an, dnz, onz); 1427 PetscCall(PetscLLCondensedCreate(Crmax, aN, &lnk, &lnkbt)); 1428 for (i = 0; i < pn; i++) { /* for each local row of C */ 1429 /* add C_loc into C */ 1430 nzi = c_loc->i[i + 1] - c_loc->i[i]; 1431 Jptr = c_loc->j + c_loc->i[i]; 1432 PetscCall(PetscLLCondensedAddSorted(nzi, Jptr, lnk, lnkbt)); 1433 1434 /* add received col data into lnk */ 1435 for (k = 0; k < nrecv; k++) { /* k-th received message */ 1436 if (i == *nextrow[k]) { /* i-th row */ 1437 nzi = *(nextci[k] + 1) - *nextci[k]; 1438 Jptr = buf_rj[k] + *nextci[k]; 1439 PetscCall(PetscLLCondensedAddSorted(nzi, Jptr, lnk, lnkbt)); 1440 nextrow[k]++; 1441 nextci[k]++; 1442 } 1443 } 1444 1445 /* add missing diagonal entry */ 1446 if (C->force_diagonals) { 1447 k = i + owners[rank]; /* column index */ 1448 PetscCall(PetscLLCondensedAddSorted(1, &k, lnk, lnkbt)); 1449 } 1450 1451 nzi = lnk[0]; 1452 1453 /* copy data into free space, then initialize lnk */ 1454 PetscCall(PetscLLCondensedClean(aN, nzi, current_space->array, lnk, lnkbt)); 1455 PetscCall(MatPreallocateSet(i + owners[rank], nzi, current_space->array, dnz, onz)); 1456 } 1457 PetscCall(PetscFree3(buf_ri_k, nextrow, nextci)); 1458 PetscCall(PetscLLDestroy(lnk, lnkbt)); 1459 PetscCall(PetscFreeSpaceDestroy(free_space)); 1460 1461 /* local sizes and preallocation */ 1462 PetscCall(MatSetSizes(C, pn, an, PETSC_DETERMINE, PETSC_DETERMINE)); 1463 PetscCall(PetscLayoutSetBlockSize(C->rmap, P->cmap->bs)); 1464 PetscCall(PetscLayoutSetBlockSize(C->cmap, A->cmap->bs)); 1465 PetscCall(MatMPIAIJSetPreallocation(C, 0, dnz, 0, onz)); 1466 MatPreallocateEnd(dnz, onz); 1467 1468 /* add C_loc and C_oth to C */ 1469 PetscCall(MatGetOwnershipRange(C, &rstart, NULL)); 1470 for (i = 0; i < pn; i++) { 1471 ncols = c_loc->i[i + 1] - c_loc->i[i]; 1472 cols = c_loc->j + c_loc->i[i]; 1473 row = rstart + i; 1474 PetscCall(MatSetValues(C, 1, (const PetscInt *)&row, ncols, (const PetscInt *)cols, NULL, INSERT_VALUES)); 1475 1476 if (C->force_diagonals) PetscCall(MatSetValues(C, 1, (const PetscInt *)&row, 1, (const PetscInt *)&row, NULL, INSERT_VALUES)); 1477 } 1478 for (i = 0; i < con; i++) { 1479 ncols = c_oth->i[i + 1] - c_oth->i[i]; 1480 cols = c_oth->j + c_oth->i[i]; 1481 row = prmap[i]; 1482 PetscCall(MatSetValues(C, 1, (const PetscInt *)&row, ncols, (const PetscInt *)cols, NULL, INSERT_VALUES)); 1483 } 1484 PetscCall(MatAssemblyBegin(C, MAT_FINAL_ASSEMBLY)); 1485 PetscCall(MatAssemblyEnd(C, MAT_FINAL_ASSEMBLY)); 1486 PetscCall(MatSetOption(C, MAT_NEW_NONZERO_LOCATION_ERR, PETSC_TRUE)); 1487 1488 /* members in merge */ 1489 PetscCall(PetscFree(id_r)); 1490 PetscCall(PetscFree(len_r)); 1491 PetscCall(PetscFree(buf_ri[0])); 1492 PetscCall(PetscFree(buf_ri)); 1493 PetscCall(PetscFree(buf_rj[0])); 1494 PetscCall(PetscFree(buf_rj)); 1495 PetscCall(PetscLayoutDestroy(&rowmap)); 1496 1497 /* attach the supporting struct to C for reuse */ 1498 C->product->data = ptap; 1499 C->product->destroy = MatProductCtxDestroy_MPIAIJ_PtAP; 1500 PetscFunctionReturn(PETSC_SUCCESS); 1501 } 1502 1503 PetscErrorCode MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ_nonscalable(Mat P, Mat A, Mat C) 1504 { 1505 Mat_MPIAIJ *p = (Mat_MPIAIJ *)P->data; 1506 Mat_SeqAIJ *c_seq; 1507 MatProductCtx_APMPI *ptap; 1508 Mat A_loc, C_loc, C_oth; 1509 PetscInt i, rstart, rend, cm, ncols, row; 1510 const PetscInt *cols; 1511 const PetscScalar *vals; 1512 1513 PetscFunctionBegin; 1514 MatCheckProduct(C, 3); 1515 ptap = (MatProductCtx_APMPI *)C->product->data; 1516 PetscCheck(ptap, PetscObjectComm((PetscObject)C), PETSC_ERR_ARG_WRONGSTATE, "PtAP cannot be computed. Missing data"); 1517 PetscCheck(ptap->A_loc, PetscObjectComm((PetscObject)C), PETSC_ERR_ARG_WRONGSTATE, "PtA cannot be reused. Do not call MatProductClear()"); 1518 PetscCall(MatZeroEntries(C)); 1519 1520 /* These matrices are obtained in MatTransposeMatMultSymbolic() */ 1521 /* 1) get R = Pd^T, Ro = Po^T */ 1522 PetscCall(MatTransposeSetPrecursor(p->A, ptap->Rd)); 1523 PetscCall(MatTranspose(p->A, MAT_REUSE_MATRIX, &ptap->Rd)); 1524 PetscCall(MatTransposeSetPrecursor(p->B, ptap->Ro)); 1525 PetscCall(MatTranspose(p->B, MAT_REUSE_MATRIX, &ptap->Ro)); 1526 1527 /* 2) compute numeric A_loc */ 1528 PetscCall(MatMPIAIJGetLocalMat(A, MAT_REUSE_MATRIX, &ptap->A_loc)); 1529 1530 /* 3) C_loc = Rd*A_loc, C_oth = Ro*A_loc */ 1531 A_loc = ptap->A_loc; 1532 PetscCall(ptap->C_loc->ops->matmultnumeric(ptap->Rd, A_loc, ptap->C_loc)); 1533 PetscCall(ptap->C_oth->ops->matmultnumeric(ptap->Ro, A_loc, ptap->C_oth)); 1534 C_loc = ptap->C_loc; 1535 C_oth = ptap->C_oth; 1536 1537 /* add C_loc and C_oth to C */ 1538 PetscCall(MatGetOwnershipRange(C, &rstart, &rend)); 1539 1540 /* C_loc -> C */ 1541 cm = C_loc->rmap->N; 1542 c_seq = (Mat_SeqAIJ *)C_loc->data; 1543 cols = c_seq->j; 1544 vals = c_seq->a; 1545 for (i = 0; i < cm; i++) { 1546 ncols = c_seq->i[i + 1] - c_seq->i[i]; 1547 row = rstart + i; 1548 PetscCall(MatSetValues(C, 1, &row, ncols, cols, vals, ADD_VALUES)); 1549 cols += ncols; 1550 vals += ncols; 1551 } 1552 1553 /* Co -> C, off-processor part */ 1554 cm = C_oth->rmap->N; 1555 c_seq = (Mat_SeqAIJ *)C_oth->data; 1556 cols = c_seq->j; 1557 vals = c_seq->a; 1558 for (i = 0; i < cm; i++) { 1559 ncols = c_seq->i[i + 1] - c_seq->i[i]; 1560 row = p->garray[i]; 1561 PetscCall(MatSetValues(C, 1, &row, ncols, cols, vals, ADD_VALUES)); 1562 cols += ncols; 1563 vals += ncols; 1564 } 1565 PetscCall(MatAssemblyBegin(C, MAT_FINAL_ASSEMBLY)); 1566 PetscCall(MatAssemblyEnd(C, MAT_FINAL_ASSEMBLY)); 1567 PetscCall(MatSetOption(C, MAT_NEW_NONZERO_LOCATION_ERR, PETSC_TRUE)); 1568 PetscFunctionReturn(PETSC_SUCCESS); 1569 } 1570 1571 PetscErrorCode MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ(Mat P, Mat A, Mat C) 1572 { 1573 MatMergeSeqsToMPI *merge; 1574 Mat_MPIAIJ *p = (Mat_MPIAIJ *)P->data; 1575 Mat_SeqAIJ *pd = (Mat_SeqAIJ *)p->A->data, *po = (Mat_SeqAIJ *)p->B->data; 1576 MatProductCtx_APMPI *ap; 1577 PetscInt *adj; 1578 PetscInt i, j, k, anz, pnz, row, *cj, nexta; 1579 MatScalar *ada, *ca, valtmp; 1580 PetscInt am = A->rmap->n, cm = C->rmap->n, pon = (p->B)->cmap->n; 1581 MPI_Comm comm; 1582 PetscMPIInt size, rank, taga, *len_s, proc; 1583 PetscInt *owners, nrows, **buf_ri_k, **nextrow, **nextci; 1584 PetscInt **buf_ri, **buf_rj; 1585 PetscInt cnz = 0, *bj_i, *bi, *bj, bnz, nextcj; /* bi,bj,ba: local array of C(mpi mat) */ 1586 MPI_Request *s_waits, *r_waits; 1587 MPI_Status *status; 1588 MatScalar **abuf_r, *ba_i, *pA, *coa, *ba; 1589 const PetscScalar *dummy; 1590 PetscInt *ai, *aj, *coi, *coj, *poJ, *pdJ; 1591 Mat A_loc; 1592 Mat_SeqAIJ *a_loc; 1593 1594 PetscFunctionBegin; 1595 MatCheckProduct(C, 3); 1596 ap = (MatProductCtx_APMPI *)C->product->data; 1597 PetscCheck(ap, PetscObjectComm((PetscObject)C), PETSC_ERR_ARG_WRONGSTATE, "PtA cannot be computed. Missing data"); 1598 PetscCheck(ap->A_loc, PetscObjectComm((PetscObject)C), PETSC_ERR_ARG_WRONGSTATE, "PtA cannot be reused. Do not call MatProductClear()"); 1599 PetscCall(PetscObjectGetComm((PetscObject)C, &comm)); 1600 PetscCallMPI(MPI_Comm_size(comm, &size)); 1601 PetscCallMPI(MPI_Comm_rank(comm, &rank)); 1602 1603 merge = ap->merge; 1604 1605 /* 2) compute numeric C_seq = P_loc^T*A_loc */ 1606 /* get data from symbolic products */ 1607 coi = merge->coi; 1608 coj = merge->coj; 1609 PetscCall(PetscCalloc1(coi[pon], &coa)); 1610 bi = merge->bi; 1611 bj = merge->bj; 1612 owners = merge->rowmap->range; 1613 PetscCall(PetscCalloc1(bi[cm], &ba)); 1614 1615 /* get A_loc by taking all local rows of A */ 1616 A_loc = ap->A_loc; 1617 PetscCall(MatMPIAIJGetLocalMat(A, MAT_REUSE_MATRIX, &A_loc)); 1618 a_loc = (Mat_SeqAIJ *)A_loc->data; 1619 ai = a_loc->i; 1620 aj = a_loc->j; 1621 1622 /* trigger copy to CPU */ 1623 PetscCall(MatSeqAIJGetArrayRead(p->A, &dummy)); 1624 PetscCall(MatSeqAIJRestoreArrayRead(p->A, &dummy)); 1625 PetscCall(MatSeqAIJGetArrayRead(p->B, &dummy)); 1626 PetscCall(MatSeqAIJRestoreArrayRead(p->B, &dummy)); 1627 for (i = 0; i < am; i++) { 1628 anz = ai[i + 1] - ai[i]; 1629 adj = aj + ai[i]; 1630 ada = a_loc->a + ai[i]; 1631 1632 /* 2-b) Compute Cseq = P_loc[i,:]^T*A[i,:] using outer product */ 1633 /* put the value into Co=(p->B)^T*A (off-diagonal part, send to others) */ 1634 pnz = po->i[i + 1] - po->i[i]; 1635 poJ = po->j + po->i[i]; 1636 pA = po->a + po->i[i]; 1637 for (j = 0; j < pnz; j++) { 1638 row = poJ[j]; 1639 cj = coj + coi[row]; 1640 ca = coa + coi[row]; 1641 /* perform sparse axpy */ 1642 nexta = 0; 1643 valtmp = pA[j]; 1644 for (k = 0; nexta < anz; k++) { 1645 if (cj[k] == adj[nexta]) { 1646 ca[k] += valtmp * ada[nexta]; 1647 nexta++; 1648 } 1649 } 1650 PetscCall(PetscLogFlops(2.0 * anz)); 1651 } 1652 1653 /* put the value into Cd (diagonal part) */ 1654 pnz = pd->i[i + 1] - pd->i[i]; 1655 pdJ = pd->j + pd->i[i]; 1656 pA = pd->a + pd->i[i]; 1657 for (j = 0; j < pnz; j++) { 1658 row = pdJ[j]; 1659 cj = bj + bi[row]; 1660 ca = ba + bi[row]; 1661 /* perform sparse axpy */ 1662 nexta = 0; 1663 valtmp = pA[j]; 1664 for (k = 0; nexta < anz; k++) { 1665 if (cj[k] == adj[nexta]) { 1666 ca[k] += valtmp * ada[nexta]; 1667 nexta++; 1668 } 1669 } 1670 PetscCall(PetscLogFlops(2.0 * anz)); 1671 } 1672 } 1673 1674 /* 3) send and recv matrix values coa */ 1675 buf_ri = merge->buf_ri; 1676 buf_rj = merge->buf_rj; 1677 len_s = merge->len_s; 1678 PetscCall(PetscCommGetNewTag(comm, &taga)); 1679 PetscCall(PetscPostIrecvScalar(comm, taga, merge->nrecv, merge->id_r, merge->len_r, &abuf_r, &r_waits)); 1680 1681 PetscCall(PetscMalloc2(merge->nsend, &s_waits, size, &status)); 1682 for (proc = 0, k = 0; proc < size; proc++) { 1683 if (!len_s[proc]) continue; 1684 i = merge->owners_co[proc]; 1685 PetscCallMPI(MPIU_Isend(coa + coi[i], len_s[proc], MPIU_MATSCALAR, proc, taga, comm, s_waits + k)); 1686 k++; 1687 } 1688 if (merge->nrecv) PetscCallMPI(MPI_Waitall(merge->nrecv, r_waits, status)); 1689 if (merge->nsend) PetscCallMPI(MPI_Waitall(merge->nsend, s_waits, status)); 1690 1691 PetscCall(PetscFree2(s_waits, status)); 1692 PetscCall(PetscFree(r_waits)); 1693 PetscCall(PetscFree(coa)); 1694 1695 /* 4) insert local Cseq and received values into Cmpi */ 1696 PetscCall(PetscMalloc3(merge->nrecv, &buf_ri_k, merge->nrecv, &nextrow, merge->nrecv, &nextci)); 1697 for (k = 0; k < merge->nrecv; k++) { 1698 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 1699 nrows = *buf_ri_k[k]; 1700 nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */ 1701 nextci[k] = buf_ri_k[k] + (nrows + 1); /* points to the next i-structure of k-th recved i-structure */ 1702 } 1703 1704 for (i = 0; i < cm; i++) { 1705 row = owners[rank] + i; /* global row index of C_seq */ 1706 bj_i = bj + bi[i]; /* col indices of the i-th row of C */ 1707 ba_i = ba + bi[i]; 1708 bnz = bi[i + 1] - bi[i]; 1709 /* add received vals into ba */ 1710 for (k = 0; k < merge->nrecv; k++) { /* k-th received message */ 1711 /* i-th row */ 1712 if (i == *nextrow[k]) { 1713 cnz = *(nextci[k] + 1) - *nextci[k]; 1714 cj = buf_rj[k] + *nextci[k]; 1715 ca = abuf_r[k] + *nextci[k]; 1716 nextcj = 0; 1717 for (j = 0; nextcj < cnz; j++) { 1718 if (bj_i[j] == cj[nextcj]) { /* bcol == ccol */ 1719 ba_i[j] += ca[nextcj++]; 1720 } 1721 } 1722 nextrow[k]++; 1723 nextci[k]++; 1724 PetscCall(PetscLogFlops(2.0 * cnz)); 1725 } 1726 } 1727 PetscCall(MatSetValues(C, 1, &row, bnz, bj_i, ba_i, INSERT_VALUES)); 1728 } 1729 PetscCall(MatAssemblyBegin(C, MAT_FINAL_ASSEMBLY)); 1730 PetscCall(MatAssemblyEnd(C, MAT_FINAL_ASSEMBLY)); 1731 1732 PetscCall(PetscFree(ba)); 1733 PetscCall(PetscFree(abuf_r[0])); 1734 PetscCall(PetscFree(abuf_r)); 1735 PetscCall(PetscFree3(buf_ri_k, nextrow, nextci)); 1736 PetscFunctionReturn(PETSC_SUCCESS); 1737 } 1738 1739 PetscErrorCode MatTransposeMatMultSymbolic_MPIAIJ_MPIAIJ(Mat P, Mat A, PetscReal fill, Mat C) 1740 { 1741 Mat A_loc; 1742 MatProductCtx_APMPI *ap; 1743 PetscFreeSpaceList free_space = NULL, current_space = NULL; 1744 Mat_MPIAIJ *p = (Mat_MPIAIJ *)P->data, *a = (Mat_MPIAIJ *)A->data; 1745 PetscInt *pdti, *pdtj, *poti, *potj, *ptJ; 1746 PetscInt nnz; 1747 PetscInt *lnk, *owners_co, *coi, *coj, i, k, pnz, row; 1748 PetscInt am = A->rmap->n, pn = P->cmap->n; 1749 MPI_Comm comm; 1750 PetscMPIInt size, rank, tagi, tagj, *len_si, *len_s, *len_ri, proc; 1751 PetscInt **buf_rj, **buf_ri, **buf_ri_k; 1752 PetscInt len, *dnz, *onz, *owners; 1753 PetscInt nzi, *bi, *bj; 1754 PetscInt nrows, *buf_s, *buf_si, *buf_si_i, **nextrow, **nextci; 1755 MPI_Request *swaits, *rwaits; 1756 MPI_Status *sstatus, rstatus; 1757 MatMergeSeqsToMPI *merge; 1758 PetscInt *ai, *aj, *Jptr, anz, *prmap = p->garray, pon, nspacedouble = 0, j; 1759 PetscReal afill = 1.0, afill_tmp; 1760 PetscInt rstart = P->cmap->rstart, rmax, Armax; 1761 Mat_SeqAIJ *a_loc; 1762 PetscHMapI ta; 1763 MatType mtype; 1764 1765 PetscFunctionBegin; 1766 PetscCall(PetscObjectGetComm((PetscObject)A, &comm)); 1767 /* check if matrix local sizes are compatible */ 1768 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, 1769 A->rmap->rend, P->rmap->rstart, P->rmap->rend); 1770 1771 PetscCallMPI(MPI_Comm_size(comm, &size)); 1772 PetscCallMPI(MPI_Comm_rank(comm, &rank)); 1773 1774 /* create struct MatProductCtx_APMPI and attached it to C later */ 1775 PetscCall(PetscNew(&ap)); 1776 1777 /* get A_loc by taking all local rows of A */ 1778 PetscCall(MatMPIAIJGetLocalMat(A, MAT_INITIAL_MATRIX, &A_loc)); 1779 1780 ap->A_loc = A_loc; 1781 a_loc = (Mat_SeqAIJ *)A_loc->data; 1782 ai = a_loc->i; 1783 aj = a_loc->j; 1784 1785 /* determine symbolic Co=(p->B)^T*A - send to others */ 1786 PetscCall(MatGetSymbolicTranspose_SeqAIJ(p->A, &pdti, &pdtj)); 1787 PetscCall(MatGetSymbolicTranspose_SeqAIJ(p->B, &poti, &potj)); 1788 pon = (p->B)->cmap->n; /* total num of rows to be sent to other processors 1789 >= (num of nonzero rows of C_seq) - pn */ 1790 PetscCall(PetscMalloc1(pon + 1, &coi)); 1791 coi[0] = 0; 1792 1793 /* set initial free space to be fill*(nnz(p->B) + nnz(A)) */ 1794 nnz = PetscRealIntMultTruncate(fill, PetscIntSumTruncate(poti[pon], ai[am])); 1795 PetscCall(PetscFreeSpaceGet(nnz, &free_space)); 1796 current_space = free_space; 1797 1798 /* create and initialize a linked list */ 1799 PetscCall(PetscHMapICreateWithSize(A->cmap->n + a->B->cmap->N, &ta)); 1800 MatRowMergeMax_SeqAIJ(a_loc, am, ta); 1801 PetscCall(PetscHMapIGetSize(ta, &Armax)); 1802 1803 PetscCall(PetscLLCondensedCreate_Scalable(Armax, &lnk)); 1804 1805 for (i = 0; i < pon; i++) { 1806 pnz = poti[i + 1] - poti[i]; 1807 ptJ = potj + poti[i]; 1808 for (j = 0; j < pnz; j++) { 1809 row = ptJ[j]; /* row of A_loc == col of Pot */ 1810 anz = ai[row + 1] - ai[row]; 1811 Jptr = aj + ai[row]; 1812 /* add non-zero cols of AP into the sorted linked list lnk */ 1813 PetscCall(PetscLLCondensedAddSorted_Scalable(anz, Jptr, lnk)); 1814 } 1815 nnz = lnk[0]; 1816 1817 /* If free space is not available, double the total space in the list */ 1818 if (current_space->local_remaining < nnz) { 1819 PetscCall(PetscFreeSpaceGet(PetscIntSumTruncate(nnz, current_space->total_array_size), ¤t_space)); 1820 nspacedouble++; 1821 } 1822 1823 /* Copy data into free space, and zero out denserows */ 1824 PetscCall(PetscLLCondensedClean_Scalable(nnz, current_space->array, lnk)); 1825 1826 current_space->array += nnz; 1827 current_space->local_used += nnz; 1828 current_space->local_remaining -= nnz; 1829 1830 coi[i + 1] = coi[i] + nnz; 1831 } 1832 1833 PetscCall(PetscMalloc1(coi[pon], &coj)); 1834 PetscCall(PetscFreeSpaceContiguous(&free_space, coj)); 1835 PetscCall(PetscLLCondensedDestroy_Scalable(lnk)); /* must destroy to get a new one for C */ 1836 1837 afill_tmp = (PetscReal)coi[pon] / (poti[pon] + ai[am] + 1); 1838 if (afill_tmp > afill) afill = afill_tmp; 1839 1840 /* send j-array (coj) of Co to other processors */ 1841 /* determine row ownership */ 1842 PetscCall(PetscNew(&merge)); 1843 PetscCall(PetscLayoutCreate(comm, &merge->rowmap)); 1844 1845 merge->rowmap->n = pn; 1846 merge->rowmap->bs = 1; 1847 1848 PetscCall(PetscLayoutSetUp(merge->rowmap)); 1849 owners = merge->rowmap->range; 1850 1851 /* determine the number of messages to send, their lengths */ 1852 PetscCall(PetscCalloc1(size, &len_si)); 1853 PetscCall(PetscCalloc1(size, &merge->len_s)); 1854 1855 len_s = merge->len_s; 1856 merge->nsend = 0; 1857 1858 PetscCall(PetscMalloc1(size + 1, &owners_co)); 1859 1860 proc = 0; 1861 for (i = 0; i < pon; i++) { 1862 while (prmap[i] >= owners[proc + 1]) proc++; 1863 len_si[proc]++; /* num of rows in Co to be sent to [proc] */ 1864 len_s[proc] += coi[i + 1] - coi[i]; 1865 } 1866 1867 len = 0; /* max length of buf_si[] */ 1868 owners_co[0] = 0; 1869 for (proc = 0; proc < size; proc++) { 1870 owners_co[proc + 1] = owners_co[proc] + len_si[proc]; 1871 if (len_s[proc]) { 1872 merge->nsend++; 1873 len_si[proc] = 2 * (len_si[proc] + 1); 1874 len += len_si[proc]; 1875 } 1876 } 1877 1878 /* determine the number and length of messages to receive for coi and coj */ 1879 PetscCall(PetscGatherNumberOfMessages(comm, NULL, len_s, &merge->nrecv)); 1880 PetscCall(PetscGatherMessageLengths2(comm, merge->nsend, merge->nrecv, len_s, len_si, &merge->id_r, &merge->len_r, &len_ri)); 1881 1882 /* post the Irecv and Isend of coj */ 1883 PetscCall(PetscCommGetNewTag(comm, &tagj)); 1884 PetscCall(PetscPostIrecvInt(comm, tagj, merge->nrecv, merge->id_r, merge->len_r, &buf_rj, &rwaits)); 1885 PetscCall(PetscMalloc1(merge->nsend, &swaits)); 1886 for (proc = 0, k = 0; proc < size; proc++) { 1887 if (!len_s[proc]) continue; 1888 i = owners_co[proc]; 1889 PetscCallMPI(MPIU_Isend(coj + coi[i], len_s[proc], MPIU_INT, proc, tagj, comm, swaits + k)); 1890 k++; 1891 } 1892 1893 /* receives and sends of coj are complete */ 1894 PetscCall(PetscMalloc1(size, &sstatus)); 1895 for (i = 0; i < merge->nrecv; i++) { 1896 PETSC_UNUSED PetscMPIInt icompleted; 1897 PetscCallMPI(MPI_Waitany(merge->nrecv, rwaits, &icompleted, &rstatus)); 1898 } 1899 PetscCall(PetscFree(rwaits)); 1900 if (merge->nsend) PetscCallMPI(MPI_Waitall(merge->nsend, swaits, sstatus)); 1901 1902 /* add received column indices into table to update Armax */ 1903 /* Armax can be as large as aN if a P[row,:] is dense, see src/ksp/ksp/tutorials/ex56.c! */ 1904 for (k = 0; k < merge->nrecv; k++) { /* k-th received message */ 1905 Jptr = buf_rj[k]; 1906 for (j = 0; j < merge->len_r[k]; j++) PetscCall(PetscHMapISet(ta, *(Jptr + j) + 1, 1)); 1907 } 1908 PetscCall(PetscHMapIGetSize(ta, &Armax)); 1909 1910 /* send and recv coi */ 1911 PetscCall(PetscCommGetNewTag(comm, &tagi)); 1912 PetscCall(PetscPostIrecvInt(comm, tagi, merge->nrecv, merge->id_r, len_ri, &buf_ri, &rwaits)); 1913 PetscCall(PetscMalloc1(len, &buf_s)); 1914 buf_si = buf_s; /* points to the beginning of k-th msg to be sent */ 1915 for (proc = 0, k = 0; proc < size; proc++) { 1916 if (!len_s[proc]) continue; 1917 /* form outgoing message for i-structure: 1918 buf_si[0]: nrows to be sent 1919 [1:nrows]: row index (global) 1920 [nrows+1:2*nrows+1]: i-structure index 1921 */ 1922 nrows = len_si[proc] / 2 - 1; 1923 buf_si_i = buf_si + nrows + 1; 1924 buf_si[0] = nrows; 1925 buf_si_i[0] = 0; 1926 nrows = 0; 1927 for (i = owners_co[proc]; i < owners_co[proc + 1]; i++) { 1928 nzi = coi[i + 1] - coi[i]; 1929 buf_si_i[nrows + 1] = buf_si_i[nrows] + nzi; /* i-structure */ 1930 buf_si[nrows + 1] = prmap[i] - owners[proc]; /* local row index */ 1931 nrows++; 1932 } 1933 PetscCallMPI(MPIU_Isend(buf_si, len_si[proc], MPIU_INT, proc, tagi, comm, swaits + k)); 1934 k++; 1935 buf_si += len_si[proc]; 1936 } 1937 i = merge->nrecv; 1938 while (i--) { 1939 PETSC_UNUSED PetscMPIInt icompleted; 1940 PetscCallMPI(MPI_Waitany(merge->nrecv, rwaits, &icompleted, &rstatus)); 1941 } 1942 PetscCall(PetscFree(rwaits)); 1943 if (merge->nsend) PetscCallMPI(MPI_Waitall(merge->nsend, swaits, sstatus)); 1944 PetscCall(PetscFree(len_si)); 1945 PetscCall(PetscFree(len_ri)); 1946 PetscCall(PetscFree(swaits)); 1947 PetscCall(PetscFree(sstatus)); 1948 PetscCall(PetscFree(buf_s)); 1949 1950 /* compute the local portion of C (mpi mat) */ 1951 /* allocate bi array and free space for accumulating nonzero column info */ 1952 PetscCall(PetscMalloc1(pn + 1, &bi)); 1953 bi[0] = 0; 1954 1955 /* set initial free space to be fill*(nnz(P) + nnz(AP)) */ 1956 nnz = PetscRealIntMultTruncate(fill, PetscIntSumTruncate(pdti[pn], PetscIntSumTruncate(poti[pon], ai[am]))); 1957 PetscCall(PetscFreeSpaceGet(nnz, &free_space)); 1958 current_space = free_space; 1959 1960 PetscCall(PetscMalloc3(merge->nrecv, &buf_ri_k, merge->nrecv, &nextrow, merge->nrecv, &nextci)); 1961 for (k = 0; k < merge->nrecv; k++) { 1962 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 1963 nrows = *buf_ri_k[k]; 1964 nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */ 1965 nextci[k] = buf_ri_k[k] + (nrows + 1); /* points to the next i-structure of k-th received i-structure */ 1966 } 1967 1968 PetscCall(PetscLLCondensedCreate_Scalable(Armax, &lnk)); 1969 MatPreallocateBegin(comm, pn, A->cmap->n, dnz, onz); 1970 rmax = 0; 1971 for (i = 0; i < pn; i++) { 1972 /* add pdt[i,:]*AP into lnk */ 1973 pnz = pdti[i + 1] - pdti[i]; 1974 ptJ = pdtj + pdti[i]; 1975 for (j = 0; j < pnz; j++) { 1976 row = ptJ[j]; /* row of AP == col of Pt */ 1977 anz = ai[row + 1] - ai[row]; 1978 Jptr = aj + ai[row]; 1979 /* add non-zero cols of AP into the sorted linked list lnk */ 1980 PetscCall(PetscLLCondensedAddSorted_Scalable(anz, Jptr, lnk)); 1981 } 1982 1983 /* add received col data into lnk */ 1984 for (k = 0; k < merge->nrecv; k++) { /* k-th received message */ 1985 if (i == *nextrow[k]) { /* i-th row */ 1986 nzi = *(nextci[k] + 1) - *nextci[k]; 1987 Jptr = buf_rj[k] + *nextci[k]; 1988 PetscCall(PetscLLCondensedAddSorted_Scalable(nzi, Jptr, lnk)); 1989 nextrow[k]++; 1990 nextci[k]++; 1991 } 1992 } 1993 1994 /* add missing diagonal entry */ 1995 if (C->force_diagonals) { 1996 k = i + owners[rank]; /* column index */ 1997 PetscCall(PetscLLCondensedAddSorted_Scalable(1, &k, lnk)); 1998 } 1999 2000 nnz = lnk[0]; 2001 2002 /* if free space is not available, make more free space */ 2003 if (current_space->local_remaining < nnz) { 2004 PetscCall(PetscFreeSpaceGet(PetscIntSumTruncate(nnz, current_space->total_array_size), ¤t_space)); 2005 nspacedouble++; 2006 } 2007 /* copy data into free space, then initialize lnk */ 2008 PetscCall(PetscLLCondensedClean_Scalable(nnz, current_space->array, lnk)); 2009 PetscCall(MatPreallocateSet(i + owners[rank], nnz, current_space->array, dnz, onz)); 2010 2011 current_space->array += nnz; 2012 current_space->local_used += nnz; 2013 current_space->local_remaining -= nnz; 2014 2015 bi[i + 1] = bi[i] + nnz; 2016 if (nnz > rmax) rmax = nnz; 2017 } 2018 PetscCall(PetscFree3(buf_ri_k, nextrow, nextci)); 2019 2020 PetscCall(PetscMalloc1(bi[pn], &bj)); 2021 PetscCall(PetscFreeSpaceContiguous(&free_space, bj)); 2022 afill_tmp = (PetscReal)bi[pn] / (pdti[pn] + poti[pon] + ai[am] + 1); 2023 if (afill_tmp > afill) afill = afill_tmp; 2024 PetscCall(PetscLLCondensedDestroy_Scalable(lnk)); 2025 PetscCall(PetscHMapIDestroy(&ta)); 2026 PetscCall(MatRestoreSymbolicTranspose_SeqAIJ(p->A, &pdti, &pdtj)); 2027 PetscCall(MatRestoreSymbolicTranspose_SeqAIJ(p->B, &poti, &potj)); 2028 2029 /* create symbolic parallel matrix C - why cannot be assembled in Numeric part */ 2030 PetscCall(MatSetSizes(C, pn, A->cmap->n, PETSC_DETERMINE, PETSC_DETERMINE)); 2031 PetscCall(MatSetBlockSizes(C, P->cmap->bs, A->cmap->bs)); 2032 PetscCall(MatGetType(A, &mtype)); 2033 PetscCall(MatSetType(C, mtype)); 2034 PetscCall(MatMPIAIJSetPreallocation(C, 0, dnz, 0, onz)); 2035 MatPreallocateEnd(dnz, onz); 2036 PetscCall(MatSetBlockSize(C, 1)); 2037 PetscCall(MatSetOption(C, MAT_NO_OFF_PROC_ENTRIES, PETSC_TRUE)); 2038 for (i = 0; i < pn; i++) { 2039 row = i + rstart; 2040 nnz = bi[i + 1] - bi[i]; 2041 Jptr = bj + bi[i]; 2042 PetscCall(MatSetValues(C, 1, &row, nnz, Jptr, NULL, INSERT_VALUES)); 2043 } 2044 PetscCall(MatAssemblyBegin(C, MAT_FINAL_ASSEMBLY)); 2045 PetscCall(MatAssemblyEnd(C, MAT_FINAL_ASSEMBLY)); 2046 PetscCall(MatSetOption(C, MAT_NEW_NONZERO_LOCATION_ERR, PETSC_TRUE)); 2047 merge->bi = bi; 2048 merge->bj = bj; 2049 merge->coi = coi; 2050 merge->coj = coj; 2051 merge->buf_ri = buf_ri; 2052 merge->buf_rj = buf_rj; 2053 merge->owners_co = owners_co; 2054 2055 /* attach the supporting struct to C for reuse */ 2056 C->product->data = ap; 2057 C->product->destroy = MatProductCtxDestroy_MPIAIJ_PtAP; 2058 ap->merge = merge; 2059 2060 C->ops->mattransposemultnumeric = MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ; 2061 2062 #if defined(PETSC_USE_INFO) 2063 if (bi[pn] != 0) { 2064 PetscCall(PetscInfo(C, "Reallocs %" PetscInt_FMT "; Fill ratio: given %g needed %g.\n", nspacedouble, (double)fill, (double)afill)); 2065 PetscCall(PetscInfo(C, "Use MatTransposeMatMult(A,B,MatReuse,%g,&C) for best performance.\n", (double)afill)); 2066 } else { 2067 PetscCall(PetscInfo(C, "Empty matrix product\n")); 2068 } 2069 #endif 2070 PetscFunctionReturn(PETSC_SUCCESS); 2071 } 2072 2073 static PetscErrorCode MatProductSymbolic_AtB_MPIAIJ_MPIAIJ(Mat C) 2074 { 2075 Mat_Product *product = C->product; 2076 Mat A = product->A, B = product->B; 2077 PetscReal fill = product->fill; 2078 PetscBool flg; 2079 2080 PetscFunctionBegin; 2081 /* scalable */ 2082 PetscCall(PetscStrcmp(product->alg, "scalable", &flg)); 2083 if (flg) { 2084 PetscCall(MatTransposeMatMultSymbolic_MPIAIJ_MPIAIJ(A, B, fill, C)); 2085 goto next; 2086 } 2087 2088 /* nonscalable */ 2089 PetscCall(PetscStrcmp(product->alg, "nonscalable", &flg)); 2090 if (flg) { 2091 PetscCall(MatTransposeMatMultSymbolic_MPIAIJ_MPIAIJ_nonscalable(A, B, fill, C)); 2092 goto next; 2093 } 2094 2095 /* matmatmult */ 2096 PetscCall(PetscStrcmp(product->alg, "at*b", &flg)); 2097 if (flg) { 2098 Mat At; 2099 MatProductCtx_APMPI *ptap; 2100 2101 PetscCall(MatTranspose(A, MAT_INITIAL_MATRIX, &At)); 2102 PetscCall(MatMatMultSymbolic_MPIAIJ_MPIAIJ(At, B, fill, C)); 2103 ptap = (MatProductCtx_APMPI *)C->product->data; 2104 if (ptap) { 2105 ptap->Pt = At; 2106 C->product->destroy = MatProductCtxDestroy_MPIAIJ_PtAP; 2107 } 2108 C->ops->transposematmultnumeric = MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ_matmatmult; 2109 goto next; 2110 } 2111 2112 /* backend general code */ 2113 PetscCall(PetscStrcmp(product->alg, "backend", &flg)); 2114 if (flg) { 2115 PetscCall(MatProductSymbolic_MPIAIJBACKEND(C)); 2116 PetscFunctionReturn(PETSC_SUCCESS); 2117 } 2118 2119 SETERRQ(PETSC_COMM_SELF, PETSC_ERR_SUP, "MatProduct type is not supported"); 2120 2121 next: 2122 C->ops->productnumeric = MatProductNumeric_AtB; 2123 PetscFunctionReturn(PETSC_SUCCESS); 2124 } 2125 2126 /* Set options for MatMatMultxxx_MPIAIJ_MPIAIJ */ 2127 static PetscErrorCode MatProductSetFromOptions_MPIAIJ_AB(Mat C) 2128 { 2129 Mat_Product *product = C->product; 2130 Mat A = product->A, B = product->B; 2131 #if defined(PETSC_HAVE_HYPRE) 2132 const char *algTypes[5] = {"scalable", "nonscalable", "seqmpi", "backend", "hypre"}; 2133 PetscInt nalg = 5; 2134 #else 2135 const char *algTypes[4] = { 2136 "scalable", 2137 "nonscalable", 2138 "seqmpi", 2139 "backend", 2140 }; 2141 PetscInt nalg = 4; 2142 #endif 2143 PetscInt alg = 1; /* set nonscalable algorithm as default */ 2144 PetscBool flg; 2145 MPI_Comm comm; 2146 2147 PetscFunctionBegin; 2148 PetscCall(PetscObjectGetComm((PetscObject)C, &comm)); 2149 2150 /* Set "nonscalable" as default algorithm */ 2151 PetscCall(PetscStrcmp(C->product->alg, "default", &flg)); 2152 if (flg) { 2153 PetscCall(MatProductSetAlgorithm(C, algTypes[alg])); 2154 2155 /* Set "scalable" as default if BN and local nonzeros of A and B are large */ 2156 if (B->cmap->N > 100000) { /* may switch to scalable algorithm as default */ 2157 MatInfo Ainfo, Binfo; 2158 PetscInt nz_local; 2159 PetscBool alg_scalable_loc = PETSC_FALSE, alg_scalable; 2160 2161 PetscCall(MatGetInfo(A, MAT_LOCAL, &Ainfo)); 2162 PetscCall(MatGetInfo(B, MAT_LOCAL, &Binfo)); 2163 nz_local = (PetscInt)(Ainfo.nz_allocated + Binfo.nz_allocated); 2164 2165 if (B->cmap->N > product->fill * nz_local) alg_scalable_loc = PETSC_TRUE; 2166 PetscCallMPI(MPIU_Allreduce(&alg_scalable_loc, &alg_scalable, 1, MPI_C_BOOL, MPI_LOR, comm)); 2167 2168 if (alg_scalable) { 2169 alg = 0; /* scalable algorithm would 50% slower than nonscalable algorithm */ 2170 PetscCall(MatProductSetAlgorithm(C, algTypes[alg])); 2171 PetscCall(PetscInfo(B, "Use scalable algorithm, BN %" PetscInt_FMT ", fill*nz_allocated %g\n", B->cmap->N, (double)(product->fill * nz_local))); 2172 } 2173 } 2174 } 2175 2176 /* Get runtime option */ 2177 if (product->api_user) { 2178 PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatMatMult", "Mat"); 2179 PetscCall(PetscOptionsEList("-matmatmult_via", "Algorithmic approach", "MatMatMult", algTypes, nalg, algTypes[alg], &alg, &flg)); 2180 PetscOptionsEnd(); 2181 } else { 2182 PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatProduct_AB", "Mat"); 2183 PetscCall(PetscOptionsEList("-mat_product_algorithm", "Algorithmic approach", "MatMatMult", algTypes, nalg, algTypes[alg], &alg, &flg)); 2184 PetscOptionsEnd(); 2185 } 2186 if (flg) PetscCall(MatProductSetAlgorithm(C, algTypes[alg])); 2187 2188 C->ops->productsymbolic = MatProductSymbolic_AB_MPIAIJ_MPIAIJ; 2189 PetscFunctionReturn(PETSC_SUCCESS); 2190 } 2191 2192 static PetscErrorCode MatProductSetFromOptions_MPIAIJ_ABt(Mat C) 2193 { 2194 PetscFunctionBegin; 2195 PetscCall(MatProductSetFromOptions_MPIAIJ_AB(C)); 2196 C->ops->productsymbolic = MatProductSymbolic_ABt_MPIAIJ_MPIAIJ; 2197 PetscFunctionReturn(PETSC_SUCCESS); 2198 } 2199 2200 /* Set options for MatTransposeMatMultXXX_MPIAIJ_MPIAIJ */ 2201 static PetscErrorCode MatProductSetFromOptions_MPIAIJ_AtB(Mat C) 2202 { 2203 Mat_Product *product = C->product; 2204 Mat A = product->A, B = product->B; 2205 const char *algTypes[4] = {"scalable", "nonscalable", "at*b", "backend"}; 2206 PetscInt nalg = 4; 2207 PetscInt alg = 1; /* set default algorithm */ 2208 PetscBool flg; 2209 MPI_Comm comm; 2210 2211 PetscFunctionBegin; 2212 /* Check matrix local sizes */ 2213 PetscCall(PetscObjectGetComm((PetscObject)C, &comm)); 2214 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 ")", 2215 A->rmap->rstart, A->rmap->rend, B->rmap->rstart, B->rmap->rend); 2216 2217 /* Set default algorithm */ 2218 PetscCall(PetscStrcmp(C->product->alg, "default", &flg)); 2219 if (flg) PetscCall(MatProductSetAlgorithm(C, algTypes[alg])); 2220 2221 /* Set "scalable" as default if BN and local nonzeros of A and B are large */ 2222 if (alg && B->cmap->N > 100000) { /* may switch to scalable algorithm as default */ 2223 MatInfo Ainfo, Binfo; 2224 PetscInt nz_local; 2225 PetscBool alg_scalable_loc = PETSC_FALSE, alg_scalable; 2226 2227 PetscCall(MatGetInfo(A, MAT_LOCAL, &Ainfo)); 2228 PetscCall(MatGetInfo(B, MAT_LOCAL, &Binfo)); 2229 nz_local = (PetscInt)(Ainfo.nz_allocated + Binfo.nz_allocated); 2230 2231 if (B->cmap->N > product->fill * nz_local) alg_scalable_loc = PETSC_TRUE; 2232 PetscCallMPI(MPIU_Allreduce(&alg_scalable_loc, &alg_scalable, 1, MPI_C_BOOL, MPI_LOR, comm)); 2233 2234 if (alg_scalable) { 2235 alg = 0; /* scalable algorithm would 50% slower than nonscalable algorithm */ 2236 PetscCall(MatProductSetAlgorithm(C, algTypes[alg])); 2237 PetscCall(PetscInfo(B, "Use scalable algorithm, BN %" PetscInt_FMT ", fill*nz_allocated %g\n", B->cmap->N, (double)(product->fill * nz_local))); 2238 } 2239 } 2240 2241 /* Get runtime option */ 2242 if (product->api_user) { 2243 PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatTransposeMatMult", "Mat"); 2244 PetscCall(PetscOptionsEList("-mattransposematmult_via", "Algorithmic approach", "MatTransposeMatMult", algTypes, nalg, algTypes[alg], &alg, &flg)); 2245 PetscOptionsEnd(); 2246 } else { 2247 PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatProduct_AtB", "Mat"); 2248 PetscCall(PetscOptionsEList("-mat_product_algorithm", "Algorithmic approach", "MatTransposeMatMult", algTypes, nalg, algTypes[alg], &alg, &flg)); 2249 PetscOptionsEnd(); 2250 } 2251 if (flg) PetscCall(MatProductSetAlgorithm(C, algTypes[alg])); 2252 2253 C->ops->productsymbolic = MatProductSymbolic_AtB_MPIAIJ_MPIAIJ; 2254 PetscFunctionReturn(PETSC_SUCCESS); 2255 } 2256 2257 static PetscErrorCode MatProductSetFromOptions_MPIAIJ_PtAP(Mat C) 2258 { 2259 Mat_Product *product = C->product; 2260 Mat A = product->A, P = product->B; 2261 MPI_Comm comm; 2262 PetscBool flg; 2263 PetscInt alg = 1; /* set default algorithm */ 2264 #if !defined(PETSC_HAVE_HYPRE) 2265 const char *algTypes[5] = {"scalable", "nonscalable", "allatonce", "allatonce_merged", "backend"}; 2266 PetscInt nalg = 5; 2267 #else 2268 const char *algTypes[6] = {"scalable", "nonscalable", "allatonce", "allatonce_merged", "backend", "hypre"}; 2269 PetscInt nalg = 6; 2270 #endif 2271 PetscInt pN = P->cmap->N; 2272 2273 PetscFunctionBegin; 2274 /* Check matrix local sizes */ 2275 PetscCall(PetscObjectGetComm((PetscObject)C, &comm)); 2276 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 ")", 2277 A->rmap->rstart, A->rmap->rend, P->rmap->rstart, P->rmap->rend); 2278 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 ")", 2279 A->cmap->rstart, A->cmap->rend, P->rmap->rstart, P->rmap->rend); 2280 2281 /* Set "nonscalable" as default algorithm */ 2282 PetscCall(PetscStrcmp(C->product->alg, "default", &flg)); 2283 if (flg) { 2284 PetscCall(MatProductSetAlgorithm(C, algTypes[alg])); 2285 2286 /* Set "scalable" as default if BN and local nonzeros of A and B are large */ 2287 if (pN > 100000) { 2288 MatInfo Ainfo, Pinfo; 2289 PetscInt nz_local; 2290 PetscBool alg_scalable_loc = PETSC_FALSE, alg_scalable; 2291 2292 PetscCall(MatGetInfo(A, MAT_LOCAL, &Ainfo)); 2293 PetscCall(MatGetInfo(P, MAT_LOCAL, &Pinfo)); 2294 nz_local = (PetscInt)(Ainfo.nz_allocated + Pinfo.nz_allocated); 2295 2296 if (pN > product->fill * nz_local) alg_scalable_loc = PETSC_TRUE; 2297 PetscCallMPI(MPIU_Allreduce(&alg_scalable_loc, &alg_scalable, 1, MPI_C_BOOL, MPI_LOR, comm)); 2298 2299 if (alg_scalable) { 2300 alg = 0; /* scalable algorithm would 50% slower than nonscalable algorithm */ 2301 PetscCall(MatProductSetAlgorithm(C, algTypes[alg])); 2302 } 2303 } 2304 } 2305 2306 /* Get runtime option */ 2307 if (product->api_user) { 2308 PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatPtAP", "Mat"); 2309 PetscCall(PetscOptionsEList("-matptap_via", "Algorithmic approach", "MatPtAP", algTypes, nalg, algTypes[alg], &alg, &flg)); 2310 PetscOptionsEnd(); 2311 } else { 2312 PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatProduct_PtAP", "Mat"); 2313 PetscCall(PetscOptionsEList("-mat_product_algorithm", "Algorithmic approach", "MatPtAP", algTypes, nalg, algTypes[alg], &alg, &flg)); 2314 PetscOptionsEnd(); 2315 } 2316 if (flg) PetscCall(MatProductSetAlgorithm(C, algTypes[alg])); 2317 2318 C->ops->productsymbolic = MatProductSymbolic_PtAP_MPIAIJ_MPIAIJ; 2319 PetscFunctionReturn(PETSC_SUCCESS); 2320 } 2321 2322 static PetscErrorCode MatProductSetFromOptions_MPIAIJ_RARt(Mat C) 2323 { 2324 Mat_Product *product = C->product; 2325 Mat A = product->A, R = product->B; 2326 2327 PetscFunctionBegin; 2328 /* Check matrix local sizes */ 2329 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, 2330 A->rmap->n, R->rmap->n, R->cmap->n); 2331 2332 C->ops->productsymbolic = MatProductSymbolic_RARt_MPIAIJ_MPIAIJ; 2333 PetscFunctionReturn(PETSC_SUCCESS); 2334 } 2335 2336 /* 2337 Set options for ABC = A*B*C = A*(B*C); ABC's algorithm must be chosen from AB's algorithm 2338 */ 2339 static PetscErrorCode MatProductSetFromOptions_MPIAIJ_ABC(Mat C) 2340 { 2341 Mat_Product *product = C->product; 2342 PetscBool flg = PETSC_FALSE; 2343 PetscInt alg = 1; /* default algorithm */ 2344 const char *algTypes[3] = {"scalable", "nonscalable", "seqmpi"}; 2345 PetscInt nalg = 3; 2346 2347 PetscFunctionBegin; 2348 /* Set default algorithm */ 2349 PetscCall(PetscStrcmp(C->product->alg, "default", &flg)); 2350 if (flg) PetscCall(MatProductSetAlgorithm(C, algTypes[alg])); 2351 2352 /* Get runtime option */ 2353 if (product->api_user) { 2354 PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatMatMatMult", "Mat"); 2355 PetscCall(PetscOptionsEList("-matmatmatmult_via", "Algorithmic approach", "MatMatMatMult", algTypes, nalg, algTypes[alg], &alg, &flg)); 2356 PetscOptionsEnd(); 2357 } else { 2358 PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatProduct_ABC", "Mat"); 2359 PetscCall(PetscOptionsEList("-mat_product_algorithm", "Algorithmic approach", "MatProduct_ABC", algTypes, nalg, algTypes[alg], &alg, &flg)); 2360 PetscOptionsEnd(); 2361 } 2362 if (flg) PetscCall(MatProductSetAlgorithm(C, algTypes[alg])); 2363 2364 C->ops->matmatmultsymbolic = MatMatMatMultSymbolic_MPIAIJ_MPIAIJ_MPIAIJ; 2365 C->ops->productsymbolic = MatProductSymbolic_ABC; 2366 PetscFunctionReturn(PETSC_SUCCESS); 2367 } 2368 2369 PETSC_INTERN PetscErrorCode MatProductSetFromOptions_MPIAIJ(Mat C) 2370 { 2371 Mat_Product *product = C->product; 2372 2373 PetscFunctionBegin; 2374 switch (product->type) { 2375 case MATPRODUCT_AB: 2376 PetscCall(MatProductSetFromOptions_MPIAIJ_AB(C)); 2377 break; 2378 case MATPRODUCT_ABt: 2379 PetscCall(MatProductSetFromOptions_MPIAIJ_ABt(C)); 2380 break; 2381 case MATPRODUCT_AtB: 2382 PetscCall(MatProductSetFromOptions_MPIAIJ_AtB(C)); 2383 break; 2384 case MATPRODUCT_PtAP: 2385 PetscCall(MatProductSetFromOptions_MPIAIJ_PtAP(C)); 2386 break; 2387 case MATPRODUCT_RARt: 2388 PetscCall(MatProductSetFromOptions_MPIAIJ_RARt(C)); 2389 break; 2390 case MATPRODUCT_ABC: 2391 PetscCall(MatProductSetFromOptions_MPIAIJ_ABC(C)); 2392 break; 2393 default: 2394 break; 2395 } 2396 PetscFunctionReturn(PETSC_SUCCESS); 2397 } 2398