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