1 2 /* 3 Defines matrix-matrix product routines for pairs of SeqAIJ matrices 4 C = A * B 5 */ 6 7 #include <../src/mat/impls/aij/seq/aij.h> /*I "petscmat.h" I*/ 8 #include <../src/mat/utils/freespace.h> 9 #include <../src/mat/utils/petscheap.h> 10 #include <petscbt.h> 11 #include <../src/mat/impls/dense/seq/dense.h> 12 13 static PetscErrorCode MatMatMultSymbolic_SeqAIJ_SeqAIJ_LLCondensed(Mat,Mat,PetscReal,Mat*); 14 15 #undef __FUNCT__ 16 #define __FUNCT__ "MatMatMult_SeqAIJ_SeqAIJ" 17 PetscErrorCode MatMatMult_SeqAIJ_SeqAIJ(Mat A,Mat B,MatReuse scall,PetscReal fill,Mat *C) 18 { 19 PetscErrorCode ierr; 20 PetscBool scalable=PETSC_FALSE,scalable_fast=PETSC_FALSE,heap = PETSC_FALSE,btheap = PETSC_FALSE,llcondensed = PETSC_FALSE; 21 22 PetscFunctionBegin; 23 if (scall == MAT_INITIAL_MATRIX) { 24 ierr = PetscObjectOptionsBegin((PetscObject)A);CHKERRQ(ierr); 25 ierr = PetscOptionsBool("-matmatmult_scalable","Use a scalable but slower C=A*B","",scalable,&scalable,NULL);CHKERRQ(ierr); 26 ierr = PetscOptionsBool("-matmatmult_scalable_fast","Use a scalable but slower C=A*B","",scalable_fast,&scalable_fast,NULL);CHKERRQ(ierr); 27 ierr = PetscOptionsBool("-matmatmult_heap","Use heap implementation of symbolic factorization C=A*B","",heap,&heap,NULL);CHKERRQ(ierr); 28 ierr = PetscOptionsBool("-matmatmult_btheap","Use btheap implementation of symbolic factorization C=A*B","",btheap,&btheap,NULL);CHKERRQ(ierr); 29 ierr = PetscOptionsBool("-matmatmult_llcondensed","Use LLCondensed to for symbolic C=A*B","",llcondensed,&llcondensed,NULL);CHKERRQ(ierr); 30 ierr = PetscOptionsEnd();CHKERRQ(ierr); 31 ierr = PetscLogEventBegin(MAT_MatMultSymbolic,A,B,0,0);CHKERRQ(ierr); 32 if (scalable_fast) { 33 ierr = MatMatMultSymbolic_SeqAIJ_SeqAIJ_Scalable_fast(A,B,fill,C);CHKERRQ(ierr); 34 } else if (scalable) { 35 ierr = MatMatMultSymbolic_SeqAIJ_SeqAIJ_Scalable(A,B,fill,C);CHKERRQ(ierr); 36 } else if (heap) { 37 ierr = MatMatMultSymbolic_SeqAIJ_SeqAIJ_Heap(A,B,fill,C);CHKERRQ(ierr); 38 } else if (btheap) { 39 ierr = MatMatMultSymbolic_SeqAIJ_SeqAIJ_BTHeap(A,B,fill,C);CHKERRQ(ierr); 40 } else if (llcondensed) { 41 ierr = MatMatMultSymbolic_SeqAIJ_SeqAIJ_LLCondensed(A,B,fill,C);CHKERRQ(ierr); 42 } else { 43 ierr = MatMatMultSymbolic_SeqAIJ_SeqAIJ(A,B,fill,C);CHKERRQ(ierr); 44 } 45 ierr = PetscLogEventEnd(MAT_MatMultSymbolic,A,B,0,0);CHKERRQ(ierr); 46 } 47 48 ierr = PetscLogEventBegin(MAT_MatMultNumeric,A,B,0,0);CHKERRQ(ierr); 49 ierr = (*(*C)->ops->matmultnumeric)(A,B,*C);CHKERRQ(ierr); 50 ierr = PetscLogEventEnd(MAT_MatMultNumeric,A,B,0,0);CHKERRQ(ierr); 51 PetscFunctionReturn(0); 52 } 53 54 #undef __FUNCT__ 55 #define __FUNCT__ "MatMatMultSymbolic_SeqAIJ_SeqAIJ_LLCondensed" 56 static PetscErrorCode MatMatMultSymbolic_SeqAIJ_SeqAIJ_LLCondensed(Mat A,Mat B,PetscReal fill,Mat *C) 57 { 58 PetscErrorCode ierr; 59 Mat_SeqAIJ *a =(Mat_SeqAIJ*)A->data,*b=(Mat_SeqAIJ*)B->data,*c; 60 PetscInt *ai=a->i,*bi=b->i,*ci,*cj; 61 PetscInt am =A->rmap->N,bn=B->cmap->N,bm=B->rmap->N; 62 PetscReal afill; 63 PetscInt i,j,anzi,brow,bnzj,cnzi,*bj,*aj,nlnk_max,*lnk,ndouble=0; 64 PetscBT lnkbt; 65 PetscFreeSpaceList free_space=NULL,current_space=NULL; 66 67 PetscFunctionBegin; 68 /* Get ci and cj */ 69 /*---------------*/ 70 /* Allocate ci array, arrays for fill computation and */ 71 /* free space for accumulating nonzero column info */ 72 ierr = PetscMalloc(((am+1)+1)*sizeof(PetscInt),&ci);CHKERRQ(ierr); 73 ci[0] = 0; 74 75 /* create and initialize a linked list */ 76 nlnk_max = a->rmax*b->rmax; 77 if (!nlnk_max || nlnk_max > bn) nlnk_max = bn; 78 ierr = PetscLLCondensedCreate(nlnk_max,bn,&lnk,&lnkbt);CHKERRQ(ierr); 79 80 /* Initial FreeSpace size is fill*(nnz(A)+nnz(B)) */ 81 ierr = PetscFreeSpaceGet((PetscInt)(fill*(ai[am]+bi[bm])),&free_space);CHKERRQ(ierr); 82 83 current_space = free_space; 84 85 /* Determine ci and cj */ 86 for (i=0; i<am; i++) { 87 anzi = ai[i+1] - ai[i]; 88 aj = a->j + ai[i]; 89 for (j=0; j<anzi; j++) { 90 brow = aj[j]; 91 bnzj = bi[brow+1] - bi[brow]; 92 bj = b->j + bi[brow]; 93 /* add non-zero cols of B into the sorted linked list lnk */ 94 ierr = PetscLLCondensedAddSorted(bnzj,bj,lnk,lnkbt);CHKERRQ(ierr); 95 } 96 cnzi = lnk[0]; 97 98 /* If free space is not available, make more free space */ 99 /* Double the amount of total space in the list */ 100 if (current_space->local_remaining<cnzi) { 101 ierr = PetscFreeSpaceGet(cnzi+current_space->total_array_size,¤t_space);CHKERRQ(ierr); 102 ndouble++; 103 } 104 105 /* Copy data into free space, then initialize lnk */ 106 ierr = PetscLLCondensedClean(bn,cnzi,current_space->array,lnk,lnkbt);CHKERRQ(ierr); 107 108 current_space->array += cnzi; 109 current_space->local_used += cnzi; 110 current_space->local_remaining -= cnzi; 111 112 ci[i+1] = ci[i] + cnzi; 113 } 114 115 /* Column indices are in the list of free space */ 116 /* Allocate space for cj, initialize cj, and */ 117 /* destroy list of free space and other temporary array(s) */ 118 ierr = PetscMalloc((ci[am]+1)*sizeof(PetscInt),&cj);CHKERRQ(ierr); 119 ierr = PetscFreeSpaceContiguous(&free_space,cj);CHKERRQ(ierr); 120 ierr = PetscLLCondensedDestroy(lnk,lnkbt);CHKERRQ(ierr); 121 122 /* put together the new symbolic matrix */ 123 ierr = MatCreateSeqAIJWithArrays(PetscObjectComm((PetscObject)A),am,bn,ci,cj,NULL,C);CHKERRQ(ierr); 124 125 (*C)->rmap->bs = A->rmap->bs; 126 (*C)->cmap->bs = B->cmap->bs; 127 128 /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */ 129 /* These are PETSc arrays, so change flags so arrays can be deleted by PETSc */ 130 c = (Mat_SeqAIJ*)((*C)->data); 131 c->free_a = PETSC_FALSE; 132 c->free_ij = PETSC_TRUE; 133 c->nonew = 0; 134 (*C)->ops->matmultnumeric = MatMatMultNumeric_SeqAIJ_SeqAIJ; /* fast, needs non-scalable O(bn) array 'abdense' */ 135 136 /* set MatInfo */ 137 afill = (PetscReal)ci[am]/(ai[am]+bi[bm]) + 1.e-5; 138 if (afill < 1.0) afill = 1.0; 139 c->maxnz = ci[am]; 140 c->nz = ci[am]; 141 (*C)->info.mallocs = ndouble; 142 (*C)->info.fill_ratio_given = fill; 143 (*C)->info.fill_ratio_needed = afill; 144 145 #if defined(PETSC_USE_INFO) 146 if (ci[am]) { 147 ierr = PetscInfo3((*C),"Reallocs %D; Fill ratio: given %G needed %G.\n",ndouble,fill,afill);CHKERRQ(ierr); 148 ierr = PetscInfo1((*C),"Use MatMatMult(A,B,MatReuse,%G,&C) for best performance.;\n",afill);CHKERRQ(ierr); 149 } else { 150 ierr = PetscInfo((*C),"Empty matrix product\n");CHKERRQ(ierr); 151 } 152 #endif 153 PetscFunctionReturn(0); 154 } 155 156 #undef __FUNCT__ 157 #define __FUNCT__ "MatMatMultNumeric_SeqAIJ_SeqAIJ" 158 PetscErrorCode MatMatMultNumeric_SeqAIJ_SeqAIJ(Mat A,Mat B,Mat C) 159 { 160 PetscErrorCode ierr; 161 PetscLogDouble flops=0.0; 162 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 163 Mat_SeqAIJ *b = (Mat_SeqAIJ*)B->data; 164 Mat_SeqAIJ *c = (Mat_SeqAIJ*)C->data; 165 PetscInt *ai =a->i,*aj=a->j,*bi=b->i,*bj=b->j,*bjj,*ci=c->i,*cj=c->j; 166 PetscInt am =A->rmap->n,cm=C->rmap->n; 167 PetscInt i,j,k,anzi,bnzi,cnzi,brow; 168 PetscScalar *aa=a->a,*ba=b->a,*baj,*ca,valtmp; 169 PetscScalar *ab_dense; 170 171 PetscFunctionBegin; 172 /* printf("MatMatMultNumeric_SeqAIJ_SeqAIJ...ca %p\n",c->a); */ 173 if (!c->a) { /* first call of MatMatMultNumeric_SeqAIJ_SeqAIJ, allocate ca and matmult_abdense */ 174 ierr = PetscMalloc((ci[cm]+1)*sizeof(MatScalar),&ca);CHKERRQ(ierr); 175 c->a = ca; 176 c->free_a = PETSC_TRUE; 177 178 ierr = PetscMalloc(B->cmap->N*sizeof(PetscScalar),&ab_dense);CHKERRQ(ierr); 179 ierr = PetscMemzero(ab_dense,B->cmap->N*sizeof(PetscScalar));CHKERRQ(ierr); 180 181 c->matmult_abdense = ab_dense; 182 } else { 183 ca = c->a; 184 ab_dense = c->matmult_abdense; 185 } 186 187 /* clean old values in C */ 188 ierr = PetscMemzero(ca,ci[cm]*sizeof(MatScalar));CHKERRQ(ierr); 189 /* Traverse A row-wise. */ 190 /* Build the ith row in C by summing over nonzero columns in A, */ 191 /* the rows of B corresponding to nonzeros of A. */ 192 for (i=0; i<am; i++) { 193 anzi = ai[i+1] - ai[i]; 194 for (j=0; j<anzi; j++) { 195 brow = aj[j]; 196 bnzi = bi[brow+1] - bi[brow]; 197 bjj = bj + bi[brow]; 198 baj = ba + bi[brow]; 199 /* perform dense axpy */ 200 valtmp = aa[j]; 201 for (k=0; k<bnzi; k++) { 202 ab_dense[bjj[k]] += valtmp*baj[k]; 203 } 204 flops += 2*bnzi; 205 } 206 aj += anzi; aa += anzi; 207 208 cnzi = ci[i+1] - ci[i]; 209 for (k=0; k<cnzi; k++) { 210 ca[k] += ab_dense[cj[k]]; 211 ab_dense[cj[k]] = 0.0; /* zero ab_dense */ 212 } 213 flops += cnzi; 214 cj += cnzi; ca += cnzi; 215 } 216 ierr = MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 217 ierr = MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 218 ierr = PetscLogFlops(flops);CHKERRQ(ierr); 219 PetscFunctionReturn(0); 220 } 221 222 #undef __FUNCT__ 223 #define __FUNCT__ "MatMatMultNumeric_SeqAIJ_SeqAIJ_Scalable" 224 PetscErrorCode MatMatMultNumeric_SeqAIJ_SeqAIJ_Scalable(Mat A,Mat B,Mat C) 225 { 226 PetscErrorCode ierr; 227 PetscLogDouble flops=0.0; 228 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 229 Mat_SeqAIJ *b = (Mat_SeqAIJ*)B->data; 230 Mat_SeqAIJ *c = (Mat_SeqAIJ*)C->data; 231 PetscInt *ai = a->i,*aj=a->j,*bi=b->i,*bj=b->j,*bjj,*ci=c->i,*cj=c->j; 232 PetscInt am = A->rmap->N,cm=C->rmap->N; 233 PetscInt i,j,k,anzi,bnzi,cnzi,brow; 234 PetscScalar *aa=a->a,*ba=b->a,*baj,*ca=c->a,valtmp; 235 PetscInt nextb; 236 237 PetscFunctionBegin; 238 /* clean old values in C */ 239 ierr = PetscMemzero(ca,ci[cm]*sizeof(MatScalar));CHKERRQ(ierr); 240 /* Traverse A row-wise. */ 241 /* Build the ith row in C by summing over nonzero columns in A, */ 242 /* the rows of B corresponding to nonzeros of A. */ 243 for (i=0; i<am; i++) { 244 anzi = ai[i+1] - ai[i]; 245 cnzi = ci[i+1] - ci[i]; 246 for (j=0; j<anzi; j++) { 247 brow = aj[j]; 248 bnzi = bi[brow+1] - bi[brow]; 249 bjj = bj + bi[brow]; 250 baj = ba + bi[brow]; 251 /* perform sparse axpy */ 252 valtmp = aa[j]; 253 nextb = 0; 254 for (k=0; nextb<bnzi; k++) { 255 if (cj[k] == bjj[nextb]) { /* ccol == bcol */ 256 ca[k] += valtmp*baj[nextb++]; 257 } 258 } 259 flops += 2*bnzi; 260 } 261 aj += anzi; aa += anzi; 262 cj += cnzi; ca += cnzi; 263 } 264 265 ierr = MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 266 ierr = MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 267 ierr = PetscLogFlops(flops);CHKERRQ(ierr); 268 PetscFunctionReturn(0); 269 } 270 271 #undef __FUNCT__ 272 #define __FUNCT__ "MatMatMultSymbolic_SeqAIJ_SeqAIJ_Scalable_fast" 273 PetscErrorCode MatMatMultSymbolic_SeqAIJ_SeqAIJ_Scalable_fast(Mat A,Mat B,PetscReal fill,Mat *C) 274 { 275 PetscErrorCode ierr; 276 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data,*b=(Mat_SeqAIJ*)B->data,*c; 277 PetscInt *ai = a->i,*bi=b->i,*ci,*cj; 278 PetscInt am = A->rmap->N,bn=B->cmap->N,bm=B->rmap->N; 279 MatScalar *ca; 280 PetscReal afill; 281 PetscInt i,j,anzi,brow,bnzj,cnzi,*bj,*aj,nlnk_max,*lnk,ndouble=0; 282 PetscFreeSpaceList free_space=NULL,current_space=NULL; 283 284 PetscFunctionBegin; 285 /* Get ci and cj - same as MatMatMultSymbolic_SeqAIJ_SeqAIJ except using PetscLLxxx_fast() */ 286 /*-----------------------------------------------------------------------------------------*/ 287 /* Allocate arrays for fill computation and free space for accumulating nonzero column */ 288 ierr = PetscMalloc(((am+1)+1)*sizeof(PetscInt),&ci);CHKERRQ(ierr); 289 ci[0] = 0; 290 291 /* create and initialize a linked list */ 292 nlnk_max = a->rmax*b->rmax; 293 if (!nlnk_max || nlnk_max > bn) nlnk_max = bn; /* in case rmax is not defined for A or B */ 294 ierr = PetscLLCondensedCreate_fast(nlnk_max,&lnk);CHKERRQ(ierr); 295 296 /* Initial FreeSpace size is fill*(nnz(A)+nnz(B)) */ 297 ierr = PetscFreeSpaceGet((PetscInt)(fill*(ai[am]+bi[bm])),&free_space);CHKERRQ(ierr); 298 current_space = free_space; 299 300 /* Determine ci and cj */ 301 for (i=0; i<am; i++) { 302 anzi = ai[i+1] - ai[i]; 303 aj = a->j + ai[i]; 304 for (j=0; j<anzi; j++) { 305 brow = aj[j]; 306 bnzj = bi[brow+1] - bi[brow]; 307 bj = b->j + bi[brow]; 308 /* add non-zero cols of B into the sorted linked list lnk */ 309 ierr = PetscLLCondensedAddSorted_fast(bnzj,bj,lnk);CHKERRQ(ierr); 310 } 311 cnzi = lnk[1]; 312 313 /* If free space is not available, make more free space */ 314 /* Double the amount of total space in the list */ 315 if (current_space->local_remaining<cnzi) { 316 ierr = PetscFreeSpaceGet(cnzi+current_space->total_array_size,¤t_space);CHKERRQ(ierr); 317 ndouble++; 318 } 319 320 /* Copy data into free space, then initialize lnk */ 321 ierr = PetscLLCondensedClean_fast(cnzi,current_space->array,lnk);CHKERRQ(ierr); 322 323 current_space->array += cnzi; 324 current_space->local_used += cnzi; 325 current_space->local_remaining -= cnzi; 326 327 ci[i+1] = ci[i] + cnzi; 328 } 329 330 /* Column indices are in the list of free space */ 331 /* Allocate space for cj, initialize cj, and */ 332 /* destroy list of free space and other temporary array(s) */ 333 ierr = PetscMalloc((ci[am]+1)*sizeof(PetscInt),&cj);CHKERRQ(ierr); 334 ierr = PetscFreeSpaceContiguous(&free_space,cj);CHKERRQ(ierr); 335 ierr = PetscLLCondensedDestroy_fast(lnk);CHKERRQ(ierr); 336 337 /* Allocate space for ca */ 338 ierr = PetscMalloc((ci[am]+1)*sizeof(MatScalar),&ca);CHKERRQ(ierr); 339 ierr = PetscMemzero(ca,(ci[am]+1)*sizeof(MatScalar));CHKERRQ(ierr); 340 341 /* put together the new symbolic matrix */ 342 ierr = MatCreateSeqAIJWithArrays(PetscObjectComm((PetscObject)A),am,bn,ci,cj,ca,C);CHKERRQ(ierr); 343 344 (*C)->rmap->bs = A->rmap->bs; 345 (*C)->cmap->bs = B->cmap->bs; 346 347 /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */ 348 /* These are PETSc arrays, so change flags so arrays can be deleted by PETSc */ 349 c = (Mat_SeqAIJ*)((*C)->data); 350 c->free_a = PETSC_TRUE; 351 c->free_ij = PETSC_TRUE; 352 c->nonew = 0; 353 354 (*C)->ops->matmultnumeric = MatMatMultNumeric_SeqAIJ_SeqAIJ_Scalable; /* slower, less memory */ 355 356 /* set MatInfo */ 357 afill = (PetscReal)ci[am]/(ai[am]+bi[bm]) + 1.e-5; 358 if (afill < 1.0) afill = 1.0; 359 c->maxnz = ci[am]; 360 c->nz = ci[am]; 361 (*C)->info.mallocs = ndouble; 362 (*C)->info.fill_ratio_given = fill; 363 (*C)->info.fill_ratio_needed = afill; 364 365 #if defined(PETSC_USE_INFO) 366 if (ci[am]) { 367 ierr = PetscInfo3((*C),"Reallocs %D; Fill ratio: given %G needed %G.\n",ndouble,fill,afill);CHKERRQ(ierr); 368 ierr = PetscInfo1((*C),"Use MatMatMult(A,B,MatReuse,%G,&C) for best performance.;\n",afill);CHKERRQ(ierr); 369 } else { 370 ierr = PetscInfo((*C),"Empty matrix product\n");CHKERRQ(ierr); 371 } 372 #endif 373 PetscFunctionReturn(0); 374 } 375 376 377 #undef __FUNCT__ 378 #define __FUNCT__ "MatMatMultSymbolic_SeqAIJ_SeqAIJ_Scalable" 379 PetscErrorCode MatMatMultSymbolic_SeqAIJ_SeqAIJ_Scalable(Mat A,Mat B,PetscReal fill,Mat *C) 380 { 381 PetscErrorCode ierr; 382 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data,*b=(Mat_SeqAIJ*)B->data,*c; 383 PetscInt *ai = a->i,*bi=b->i,*ci,*cj; 384 PetscInt am = A->rmap->N,bn=B->cmap->N,bm=B->rmap->N; 385 MatScalar *ca; 386 PetscReal afill; 387 PetscInt i,j,anzi,brow,bnzj,cnzi,*bj,*aj,nlnk_max,*lnk,ndouble=0; 388 PetscFreeSpaceList free_space=NULL,current_space=NULL; 389 390 PetscFunctionBegin; 391 /* Get ci and cj - same as MatMatMultSymbolic_SeqAIJ_SeqAIJ except using PetscLLxxx_Scalalbe() */ 392 /*---------------------------------------------------------------------------------------------*/ 393 /* Allocate arrays for fill computation and free space for accumulating nonzero column */ 394 ierr = PetscMalloc(((am+1)+1)*sizeof(PetscInt),&ci);CHKERRQ(ierr); 395 ci[0] = 0; 396 397 /* create and initialize a linked list */ 398 nlnk_max = a->rmax*b->rmax; 399 if (!nlnk_max || nlnk_max > bn) nlnk_max = bn; /* in case rmax is not defined for A or B */ 400 ierr = PetscLLCondensedCreate_Scalable(nlnk_max,&lnk);CHKERRQ(ierr); 401 402 /* Initial FreeSpace size is fill*(nnz(A)+nnz(B)) */ 403 ierr = PetscFreeSpaceGet((PetscInt)(fill*(ai[am]+bi[bm])),&free_space);CHKERRQ(ierr); 404 current_space = free_space; 405 406 /* Determine ci and cj */ 407 for (i=0; i<am; i++) { 408 anzi = ai[i+1] - ai[i]; 409 aj = a->j + ai[i]; 410 for (j=0; j<anzi; j++) { 411 brow = aj[j]; 412 bnzj = bi[brow+1] - bi[brow]; 413 bj = b->j + bi[brow]; 414 /* add non-zero cols of B into the sorted linked list lnk */ 415 ierr = PetscLLCondensedAddSorted_Scalable(bnzj,bj,lnk);CHKERRQ(ierr); 416 } 417 cnzi = lnk[0]; 418 419 /* If free space is not available, make more free space */ 420 /* Double the amount of total space in the list */ 421 if (current_space->local_remaining<cnzi) { 422 ierr = PetscFreeSpaceGet(cnzi+current_space->total_array_size,¤t_space);CHKERRQ(ierr); 423 ndouble++; 424 } 425 426 /* Copy data into free space, then initialize lnk */ 427 ierr = PetscLLCondensedClean_Scalable(cnzi,current_space->array,lnk);CHKERRQ(ierr); 428 429 current_space->array += cnzi; 430 current_space->local_used += cnzi; 431 current_space->local_remaining -= cnzi; 432 433 ci[i+1] = ci[i] + cnzi; 434 } 435 436 /* Column indices are in the list of free space */ 437 /* Allocate space for cj, initialize cj, and */ 438 /* destroy list of free space and other temporary array(s) */ 439 ierr = PetscMalloc((ci[am]+1)*sizeof(PetscInt),&cj);CHKERRQ(ierr); 440 ierr = PetscFreeSpaceContiguous(&free_space,cj);CHKERRQ(ierr); 441 ierr = PetscLLCondensedDestroy_Scalable(lnk);CHKERRQ(ierr); 442 443 /* Allocate space for ca */ 444 /*-----------------------*/ 445 ierr = PetscMalloc((ci[am]+1)*sizeof(MatScalar),&ca);CHKERRQ(ierr); 446 ierr = PetscMemzero(ca,(ci[am]+1)*sizeof(MatScalar));CHKERRQ(ierr); 447 448 /* put together the new symbolic matrix */ 449 ierr = MatCreateSeqAIJWithArrays(PetscObjectComm((PetscObject)A),am,bn,ci,cj,ca,C);CHKERRQ(ierr); 450 451 (*C)->rmap->bs = A->rmap->bs; 452 (*C)->cmap->bs = B->cmap->bs; 453 454 /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */ 455 /* These are PETSc arrays, so change flags so arrays can be deleted by PETSc */ 456 c = (Mat_SeqAIJ*)((*C)->data); 457 c->free_a = PETSC_TRUE; 458 c->free_ij = PETSC_TRUE; 459 c->nonew = 0; 460 461 (*C)->ops->matmultnumeric = MatMatMultNumeric_SeqAIJ_SeqAIJ_Scalable; /* slower, less memory */ 462 463 /* set MatInfo */ 464 afill = (PetscReal)ci[am]/(ai[am]+bi[bm]) + 1.e-5; 465 if (afill < 1.0) afill = 1.0; 466 c->maxnz = ci[am]; 467 c->nz = ci[am]; 468 (*C)->info.mallocs = ndouble; 469 (*C)->info.fill_ratio_given = fill; 470 (*C)->info.fill_ratio_needed = afill; 471 472 #if defined(PETSC_USE_INFO) 473 if (ci[am]) { 474 ierr = PetscInfo3((*C),"Reallocs %D; Fill ratio: given %G needed %G.\n",ndouble,fill,afill);CHKERRQ(ierr); 475 ierr = PetscInfo1((*C),"Use MatMatMult(A,B,MatReuse,%G,&C) for best performance.;\n",afill);CHKERRQ(ierr); 476 } else { 477 ierr = PetscInfo((*C),"Empty matrix product\n");CHKERRQ(ierr); 478 } 479 #endif 480 PetscFunctionReturn(0); 481 } 482 483 #undef __FUNCT__ 484 #define __FUNCT__ "MatMatMultSymbolic_SeqAIJ_SeqAIJ_Heap" 485 PetscErrorCode MatMatMultSymbolic_SeqAIJ_SeqAIJ_Heap(Mat A,Mat B,PetscReal fill,Mat *C) 486 { 487 PetscErrorCode ierr; 488 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data,*b=(Mat_SeqAIJ*)B->data,*c; 489 const PetscInt *ai=a->i,*bi=b->i,*aj=a->j,*bj=b->j; 490 PetscInt *ci,*cj,*bb; 491 PetscInt am=A->rmap->N,bn=B->cmap->N,bm=B->rmap->N; 492 PetscReal afill; 493 PetscInt i,j,col,ndouble = 0; 494 PetscFreeSpaceList free_space=NULL,current_space=NULL; 495 PetscHeap h; 496 497 PetscFunctionBegin; 498 /* Get ci and cj - by merging sorted rows using a heap */ 499 /*---------------------------------------------------------------------------------------------*/ 500 /* Allocate arrays for fill computation and free space for accumulating nonzero column */ 501 ierr = PetscMalloc(((am+1)+1)*sizeof(PetscInt),&ci);CHKERRQ(ierr); 502 ci[0] = 0; 503 504 /* Initial FreeSpace size is fill*(nnz(A)+nnz(B)) */ 505 ierr = PetscFreeSpaceGet((PetscInt)(fill*(ai[am]+bi[bm])),&free_space);CHKERRQ(ierr); 506 current_space = free_space; 507 508 ierr = PetscHeapCreate(a->rmax,&h);CHKERRQ(ierr); 509 ierr = PetscMalloc(a->rmax*sizeof(PetscInt),&bb);CHKERRQ(ierr); 510 511 /* Determine ci and cj */ 512 for (i=0; i<am; i++) { 513 const PetscInt anzi = ai[i+1] - ai[i]; /* number of nonzeros in this row of A, this is the number of rows of B that we merge */ 514 const PetscInt *acol = aj + ai[i]; /* column indices of nonzero entries in this row */ 515 ci[i+1] = ci[i]; 516 /* Populate the min heap */ 517 for (j=0; j<anzi; j++) { 518 bb[j] = bi[acol[j]]; /* bb points at the start of the row */ 519 if (bb[j] < bi[acol[j]+1]) { /* Add if row is nonempty */ 520 ierr = PetscHeapAdd(h,j,bj[bb[j]++]);CHKERRQ(ierr); 521 } 522 } 523 /* Pick off the min element, adding it to free space */ 524 ierr = PetscHeapPop(h,&j,&col);CHKERRQ(ierr); 525 while (j >= 0) { 526 if (current_space->local_remaining < 1) { /* double the size, but don't exceed 16 MiB */ 527 ierr = PetscFreeSpaceGet(PetscMin(2*current_space->total_array_size,16 << 20),¤t_space);CHKERRQ(ierr); 528 ndouble++; 529 } 530 *(current_space->array++) = col; 531 current_space->local_used++; 532 current_space->local_remaining--; 533 ci[i+1]++; 534 535 /* stash if anything else remains in this row of B */ 536 if (bb[j] < bi[acol[j]+1]) {ierr = PetscHeapStash(h,j,bj[bb[j]++]);CHKERRQ(ierr);} 537 while (1) { /* pop and stash any other rows of B that also had an entry in this column */ 538 PetscInt j2,col2; 539 ierr = PetscHeapPeek(h,&j2,&col2);CHKERRQ(ierr); 540 if (col2 != col) break; 541 ierr = PetscHeapPop(h,&j2,&col2);CHKERRQ(ierr); 542 if (bb[j2] < bi[acol[j2]+1]) {ierr = PetscHeapStash(h,j2,bj[bb[j2]++]);CHKERRQ(ierr);} 543 } 544 /* Put any stashed elements back into the min heap */ 545 ierr = PetscHeapUnstash(h);CHKERRQ(ierr); 546 ierr = PetscHeapPop(h,&j,&col);CHKERRQ(ierr); 547 } 548 } 549 ierr = PetscFree(bb);CHKERRQ(ierr); 550 ierr = PetscHeapDestroy(&h);CHKERRQ(ierr); 551 552 /* Column indices are in the list of free space */ 553 /* Allocate space for cj, initialize cj, and */ 554 /* destroy list of free space and other temporary array(s) */ 555 ierr = PetscMalloc(ci[am]*sizeof(PetscInt),&cj);CHKERRQ(ierr); 556 ierr = PetscFreeSpaceContiguous(&free_space,cj);CHKERRQ(ierr); 557 558 /* put together the new symbolic matrix */ 559 ierr = MatCreateSeqAIJWithArrays(PetscObjectComm((PetscObject)A),am,bn,ci,cj,NULL,C);CHKERRQ(ierr); 560 561 (*C)->rmap->bs = A->rmap->bs; 562 (*C)->cmap->bs = B->cmap->bs; 563 564 /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */ 565 /* These are PETSc arrays, so change flags so arrays can be deleted by PETSc */ 566 c = (Mat_SeqAIJ*)((*C)->data); 567 c->free_a = PETSC_TRUE; 568 c->free_ij = PETSC_TRUE; 569 c->nonew = 0; 570 571 (*C)->ops->matmultnumeric = MatMatMultNumeric_SeqAIJ_SeqAIJ; 572 573 /* set MatInfo */ 574 afill = (PetscReal)ci[am]/(ai[am]+bi[bm]) + 1.e-5; 575 if (afill < 1.0) afill = 1.0; 576 c->maxnz = ci[am]; 577 c->nz = ci[am]; 578 (*C)->info.mallocs = ndouble; 579 (*C)->info.fill_ratio_given = fill; 580 (*C)->info.fill_ratio_needed = afill; 581 582 #if defined(PETSC_USE_INFO) 583 if (ci[am]) { 584 ierr = PetscInfo3((*C),"Reallocs %D; Fill ratio: given %G needed %G.\n",ndouble,fill,afill);CHKERRQ(ierr); 585 ierr = PetscInfo1((*C),"Use MatMatMult(A,B,MatReuse,%G,&C) for best performance.;\n",afill);CHKERRQ(ierr); 586 } else { 587 ierr = PetscInfo((*C),"Empty matrix product\n");CHKERRQ(ierr); 588 } 589 #endif 590 PetscFunctionReturn(0); 591 } 592 593 #undef __FUNCT__ 594 #define __FUNCT__ "MatMatMultSymbolic_SeqAIJ_SeqAIJ_BTHeap" 595 PetscErrorCode MatMatMultSymbolic_SeqAIJ_SeqAIJ_BTHeap(Mat A,Mat B,PetscReal fill,Mat *C) 596 { 597 PetscErrorCode ierr; 598 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data,*b=(Mat_SeqAIJ*)B->data,*c; 599 const PetscInt *ai = a->i,*bi=b->i,*aj=a->j,*bj=b->j; 600 PetscInt *ci,*cj,*bb; 601 PetscInt am=A->rmap->N,bn=B->cmap->N,bm=B->rmap->N; 602 PetscReal afill; 603 PetscInt i,j,col,ndouble = 0; 604 PetscFreeSpaceList free_space=NULL,current_space=NULL; 605 PetscHeap h; 606 PetscBT bt; 607 608 PetscFunctionBegin; 609 /* Get ci and cj - using a heap for the sorted rows, but use BT so that each index is only added once */ 610 /*---------------------------------------------------------------------------------------------*/ 611 /* Allocate arrays for fill computation and free space for accumulating nonzero column */ 612 ierr = PetscMalloc(((am+1)+1)*sizeof(PetscInt),&ci);CHKERRQ(ierr); 613 ci[0] = 0; 614 615 /* Initial FreeSpace size is fill*(nnz(A)+nnz(B)) */ 616 ierr = PetscFreeSpaceGet((PetscInt)(fill*(ai[am]+bi[bm])),&free_space);CHKERRQ(ierr); 617 618 current_space = free_space; 619 620 ierr = PetscHeapCreate(a->rmax,&h);CHKERRQ(ierr); 621 ierr = PetscMalloc(a->rmax*sizeof(PetscInt),&bb);CHKERRQ(ierr); 622 ierr = PetscBTCreate(bn,&bt);CHKERRQ(ierr); 623 624 /* Determine ci and cj */ 625 for (i=0; i<am; i++) { 626 const PetscInt anzi = ai[i+1] - ai[i]; /* number of nonzeros in this row of A, this is the number of rows of B that we merge */ 627 const PetscInt *acol = aj + ai[i]; /* column indices of nonzero entries in this row */ 628 const PetscInt *fptr = current_space->array; /* Save beginning of the row so we can clear the BT later */ 629 ci[i+1] = ci[i]; 630 /* Populate the min heap */ 631 for (j=0; j<anzi; j++) { 632 PetscInt brow = acol[j]; 633 for (bb[j] = bi[brow]; bb[j] < bi[brow+1]; bb[j]++) { 634 PetscInt bcol = bj[bb[j]]; 635 if (!PetscBTLookupSet(bt,bcol)) { /* new entry */ 636 ierr = PetscHeapAdd(h,j,bcol);CHKERRQ(ierr); 637 bb[j]++; 638 break; 639 } 640 } 641 } 642 /* Pick off the min element, adding it to free space */ 643 ierr = PetscHeapPop(h,&j,&col);CHKERRQ(ierr); 644 while (j >= 0) { 645 if (current_space->local_remaining < 1) { /* double the size, but don't exceed 16 MiB */ 646 fptr = NULL; /* need PetscBTMemzero */ 647 ierr = PetscFreeSpaceGet(PetscMin(2*current_space->total_array_size,16 << 20),¤t_space);CHKERRQ(ierr); 648 ndouble++; 649 } 650 *(current_space->array++) = col; 651 current_space->local_used++; 652 current_space->local_remaining--; 653 ci[i+1]++; 654 655 /* stash if anything else remains in this row of B */ 656 for (; bb[j] < bi[acol[j]+1]; bb[j]++) { 657 PetscInt bcol = bj[bb[j]]; 658 if (!PetscBTLookupSet(bt,bcol)) { /* new entry */ 659 ierr = PetscHeapAdd(h,j,bcol);CHKERRQ(ierr); 660 bb[j]++; 661 break; 662 } 663 } 664 ierr = PetscHeapPop(h,&j,&col);CHKERRQ(ierr); 665 } 666 if (fptr) { /* Clear the bits for this row */ 667 for (; fptr<current_space->array; fptr++) {ierr = PetscBTClear(bt,*fptr);CHKERRQ(ierr);} 668 } else { /* We reallocated so we don't remember (easily) how to clear only the bits we changed */ 669 ierr = PetscBTMemzero(bn,bt);CHKERRQ(ierr); 670 } 671 } 672 ierr = PetscFree(bb);CHKERRQ(ierr); 673 ierr = PetscHeapDestroy(&h);CHKERRQ(ierr); 674 ierr = PetscBTDestroy(&bt);CHKERRQ(ierr); 675 676 /* Column indices are in the list of free space */ 677 /* Allocate space for cj, initialize cj, and */ 678 /* destroy list of free space and other temporary array(s) */ 679 ierr = PetscMalloc(ci[am]*sizeof(PetscInt),&cj);CHKERRQ(ierr); 680 ierr = PetscFreeSpaceContiguous(&free_space,cj);CHKERRQ(ierr); 681 682 /* put together the new symbolic matrix */ 683 ierr = MatCreateSeqAIJWithArrays(PetscObjectComm((PetscObject)A),am,bn,ci,cj,NULL,C);CHKERRQ(ierr); 684 685 (*C)->rmap->bs = A->rmap->bs; 686 (*C)->cmap->bs = B->cmap->bs; 687 688 /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */ 689 /* These are PETSc arrays, so change flags so arrays can be deleted by PETSc */ 690 c = (Mat_SeqAIJ*)((*C)->data); 691 c->free_a = PETSC_TRUE; 692 c->free_ij = PETSC_TRUE; 693 c->nonew = 0; 694 695 (*C)->ops->matmultnumeric = MatMatMultNumeric_SeqAIJ_SeqAIJ; 696 697 /* set MatInfo */ 698 afill = (PetscReal)ci[am]/(ai[am]+bi[bm]) + 1.e-5; 699 if (afill < 1.0) afill = 1.0; 700 c->maxnz = ci[am]; 701 c->nz = ci[am]; 702 (*C)->info.mallocs = ndouble; 703 (*C)->info.fill_ratio_given = fill; 704 (*C)->info.fill_ratio_needed = afill; 705 706 #if defined(PETSC_USE_INFO) 707 if (ci[am]) { 708 ierr = PetscInfo3((*C),"Reallocs %D; Fill ratio: given %G needed %G.\n",ndouble,fill,afill);CHKERRQ(ierr); 709 ierr = PetscInfo1((*C),"Use MatMatMult(A,B,MatReuse,%G,&C) for best performance.;\n",afill);CHKERRQ(ierr); 710 } else { 711 ierr = PetscInfo((*C),"Empty matrix product\n");CHKERRQ(ierr); 712 } 713 #endif 714 PetscFunctionReturn(0); 715 } 716 717 #undef __FUNCT__ 718 #define __FUNCT__ "MatMatMultSymbolic_SeqAIJ_SeqAIJ" 719 /* concatenate unique entries and then sort */ 720 PetscErrorCode MatMatMultSymbolic_SeqAIJ_SeqAIJ(Mat A,Mat B,PetscReal fill,Mat *C) 721 { 722 PetscErrorCode ierr; 723 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data,*b=(Mat_SeqAIJ*)B->data,*c; 724 const PetscInt *ai = a->i,*bi=b->i,*aj=a->j,*bj=b->j; 725 PetscInt *ci,*cj; 726 PetscInt am=A->rmap->N,bn=B->cmap->N,bm=B->rmap->N; 727 PetscReal afill; 728 PetscInt i,j,ndouble = 0; 729 PetscSegBuffer seg,segrow; 730 char *seen; 731 732 PetscFunctionBegin; 733 ierr = PetscMalloc((am+1)*sizeof(PetscInt),&ci);CHKERRQ(ierr); 734 ci[0] = 0; 735 736 /* Initial FreeSpace size is fill*(nnz(A)+nnz(B)) */ 737 ierr = PetscSegBufferCreate(sizeof(PetscInt),(PetscInt)(fill*(ai[am]+bi[bm])),&seg);CHKERRQ(ierr); 738 ierr = PetscSegBufferCreate(sizeof(PetscInt),100,&segrow);CHKERRQ(ierr); 739 ierr = PetscMalloc(bn*sizeof(char),&seen);CHKERRQ(ierr); 740 ierr = PetscMemzero(seen,bn*sizeof(char));CHKERRQ(ierr); 741 742 /* Determine ci and cj */ 743 for (i=0; i<am; i++) { 744 const PetscInt anzi = ai[i+1] - ai[i]; /* number of nonzeros in this row of A, this is the number of rows of B that we merge */ 745 const PetscInt *acol = aj + ai[i]; /* column indices of nonzero entries in this row */ 746 PetscInt packlen = 0,*PETSC_RESTRICT crow; 747 /* Pack segrow */ 748 for (j=0; j<anzi; j++) { 749 PetscInt brow = acol[j],bjstart = bi[brow],bjend = bi[brow+1],k; 750 for (k=bjstart; k<bjend; k++) { 751 PetscInt bcol = bj[k]; 752 if (!seen[bcol]) { /* new entry */ 753 PetscInt *PETSC_RESTRICT slot; 754 ierr = PetscSegBufferGetInts(segrow,1,&slot);CHKERRQ(ierr); 755 *slot = bcol; 756 seen[bcol] = 1; 757 packlen++; 758 } 759 } 760 } 761 ierr = PetscSegBufferGetInts(seg,packlen,&crow);CHKERRQ(ierr); 762 ierr = PetscSegBufferExtractTo(segrow,crow);CHKERRQ(ierr); 763 ierr = PetscSortInt(packlen,crow);CHKERRQ(ierr); 764 ci[i+1] = ci[i] + packlen; 765 for (j=0; j<packlen; j++) seen[crow[j]] = 0; 766 } 767 ierr = PetscSegBufferDestroy(&segrow);CHKERRQ(ierr); 768 ierr = PetscFree(seen);CHKERRQ(ierr); 769 770 /* Column indices are in the segmented buffer */ 771 ierr = PetscSegBufferExtractAlloc(seg,&cj);CHKERRQ(ierr); 772 ierr = PetscSegBufferDestroy(&seg);CHKERRQ(ierr); 773 774 /* put together the new symbolic matrix */ 775 ierr = MatCreateSeqAIJWithArrays(PetscObjectComm((PetscObject)A),am,bn,ci,cj,NULL,C);CHKERRQ(ierr); 776 777 (*C)->rmap->bs = A->rmap->bs; 778 (*C)->cmap->bs = B->cmap->bs; 779 780 /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */ 781 /* These are PETSc arrays, so change flags so arrays can be deleted by PETSc */ 782 c = (Mat_SeqAIJ*)((*C)->data); 783 c->free_a = PETSC_TRUE; 784 c->free_ij = PETSC_TRUE; 785 c->nonew = 0; 786 787 (*C)->ops->matmultnumeric = MatMatMultNumeric_SeqAIJ_SeqAIJ; 788 789 /* set MatInfo */ 790 afill = (PetscReal)ci[am]/(ai[am]+bi[bm]) + 1.e-5; 791 if (afill < 1.0) afill = 1.0; 792 c->maxnz = ci[am]; 793 c->nz = ci[am]; 794 (*C)->info.mallocs = ndouble; 795 (*C)->info.fill_ratio_given = fill; 796 (*C)->info.fill_ratio_needed = afill; 797 798 #if defined(PETSC_USE_INFO) 799 if (ci[am]) { 800 ierr = PetscInfo3((*C),"Reallocs %D; Fill ratio: given %G needed %G.\n",ndouble,fill,afill);CHKERRQ(ierr); 801 ierr = PetscInfo1((*C),"Use MatMatMult(A,B,MatReuse,%G,&C) for best performance.;\n",afill);CHKERRQ(ierr); 802 } else { 803 ierr = PetscInfo((*C),"Empty matrix product\n");CHKERRQ(ierr); 804 } 805 #endif 806 PetscFunctionReturn(0); 807 } 808 809 /* This routine is not used. Should be removed! */ 810 #undef __FUNCT__ 811 #define __FUNCT__ "MatMatTransposeMult_SeqAIJ_SeqAIJ" 812 PetscErrorCode MatMatTransposeMult_SeqAIJ_SeqAIJ(Mat A,Mat B,MatReuse scall,PetscReal fill,Mat *C) 813 { 814 PetscErrorCode ierr; 815 816 PetscFunctionBegin; 817 if (scall == MAT_INITIAL_MATRIX) { 818 ierr = PetscLogEventBegin(MAT_MatTransposeMultSymbolic,A,B,0,0);CHKERRQ(ierr); 819 ierr = MatMatTransposeMultSymbolic_SeqAIJ_SeqAIJ(A,B,fill,C);CHKERRQ(ierr); 820 ierr = PetscLogEventEnd(MAT_MatTransposeMultSymbolic,A,B,0,0);CHKERRQ(ierr); 821 } 822 ierr = PetscLogEventBegin(MAT_MatTransposeMultNumeric,A,B,0,0);CHKERRQ(ierr); 823 ierr = MatMatTransposeMultNumeric_SeqAIJ_SeqAIJ(A,B,*C);CHKERRQ(ierr); 824 ierr = PetscLogEventEnd(MAT_MatTransposeMultNumeric,A,B,0,0);CHKERRQ(ierr); 825 PetscFunctionReturn(0); 826 } 827 828 #undef __FUNCT__ 829 #define __FUNCT__ "PetscContainerDestroy_Mat_MatMatTransMult" 830 PetscErrorCode PetscContainerDestroy_Mat_MatMatTransMult(void *ptr) 831 { 832 PetscErrorCode ierr; 833 Mat_MatMatTransMult *multtrans=(Mat_MatMatTransMult*)ptr; 834 835 PetscFunctionBegin; 836 ierr = MatTransposeColoringDestroy(&multtrans->matcoloring);CHKERRQ(ierr); 837 ierr = MatDestroy(&multtrans->Bt_den);CHKERRQ(ierr); 838 ierr = MatDestroy(&multtrans->ABt_den);CHKERRQ(ierr); 839 ierr = PetscFree(multtrans);CHKERRQ(ierr); 840 PetscFunctionReturn(0); 841 } 842 843 #undef __FUNCT__ 844 #define __FUNCT__ "MatDestroy_SeqAIJ_MatMatMultTrans" 845 PetscErrorCode MatDestroy_SeqAIJ_MatMatMultTrans(Mat A) 846 { 847 PetscErrorCode ierr; 848 PetscContainer container; 849 Mat_MatMatTransMult *multtrans=NULL; 850 851 PetscFunctionBegin; 852 ierr = PetscObjectQuery((PetscObject)A,"Mat_MatMatTransMult",(PetscObject*)&container);CHKERRQ(ierr); 853 if (!container) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Container does not exit"); 854 ierr = PetscContainerGetPointer(container,(void**)&multtrans);CHKERRQ(ierr); 855 856 A->ops->destroy = multtrans->destroy; 857 if (A->ops->destroy) { 858 ierr = (*A->ops->destroy)(A);CHKERRQ(ierr); 859 } 860 ierr = PetscObjectCompose((PetscObject)A,"Mat_MatMatTransMult",0);CHKERRQ(ierr); 861 PetscFunctionReturn(0); 862 } 863 864 #undef __FUNCT__ 865 #define __FUNCT__ "MatMatTransposeMultSymbolic_SeqAIJ_SeqAIJ" 866 PetscErrorCode MatMatTransposeMultSymbolic_SeqAIJ_SeqAIJ(Mat A,Mat B,PetscReal fill,Mat *C) 867 { 868 PetscErrorCode ierr; 869 Mat Bt; 870 PetscInt *bti,*btj; 871 Mat_MatMatTransMult *multtrans; 872 PetscContainer container; 873 874 PetscFunctionBegin; 875 /* create symbolic Bt */ 876 ierr = MatGetSymbolicTranspose_SeqAIJ(B,&bti,&btj);CHKERRQ(ierr); 877 ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,B->cmap->n,B->rmap->n,bti,btj,NULL,&Bt);CHKERRQ(ierr); 878 879 Bt->rmap->bs = A->cmap->bs; 880 Bt->cmap->bs = B->cmap->bs; 881 882 /* get symbolic C=A*Bt */ 883 ierr = MatMatMultSymbolic_SeqAIJ_SeqAIJ(A,Bt,fill,C);CHKERRQ(ierr); 884 885 /* create a supporting struct for reuse intermidiate dense matrices with matcoloring */ 886 ierr = PetscNew(Mat_MatMatTransMult,&multtrans);CHKERRQ(ierr); 887 888 /* attach the supporting struct to C */ 889 ierr = PetscContainerCreate(PETSC_COMM_SELF,&container);CHKERRQ(ierr); 890 ierr = PetscContainerSetPointer(container,multtrans);CHKERRQ(ierr); 891 ierr = PetscContainerSetUserDestroy(container,PetscContainerDestroy_Mat_MatMatTransMult);CHKERRQ(ierr); 892 ierr = PetscObjectCompose((PetscObject)(*C),"Mat_MatMatTransMult",(PetscObject)container);CHKERRQ(ierr); 893 ierr = PetscContainerDestroy(&container);CHKERRQ(ierr); 894 895 multtrans->usecoloring = PETSC_FALSE; 896 multtrans->destroy = (*C)->ops->destroy; 897 (*C)->ops->destroy = MatDestroy_SeqAIJ_MatMatMultTrans; 898 899 ierr = PetscOptionsGetBool(NULL,"-matmattransmult_color",&multtrans->usecoloring,NULL);CHKERRQ(ierr); 900 if (multtrans->usecoloring) { 901 /* Create MatTransposeColoring from symbolic C=A*B^T */ 902 MatTransposeColoring matcoloring; 903 ISColoring iscoloring; 904 Mat Bt_dense,C_dense; 905 906 ierr = MatGetColoring(*C,MATCOLORINGLF,&iscoloring);CHKERRQ(ierr); 907 ierr = MatTransposeColoringCreate(*C,iscoloring,&matcoloring);CHKERRQ(ierr); 908 909 multtrans->matcoloring = matcoloring; 910 911 ierr = ISColoringDestroy(&iscoloring);CHKERRQ(ierr); 912 913 /* Create Bt_dense and C_dense = A*Bt_dense */ 914 ierr = MatCreate(PETSC_COMM_SELF,&Bt_dense);CHKERRQ(ierr); 915 ierr = MatSetSizes(Bt_dense,A->cmap->n,matcoloring->ncolors,A->cmap->n,matcoloring->ncolors);CHKERRQ(ierr); 916 ierr = MatSetType(Bt_dense,MATSEQDENSE);CHKERRQ(ierr); 917 ierr = MatSeqDenseSetPreallocation(Bt_dense,NULL);CHKERRQ(ierr); 918 919 Bt_dense->assembled = PETSC_TRUE; 920 multtrans->Bt_den = Bt_dense; 921 922 ierr = MatCreate(PETSC_COMM_SELF,&C_dense);CHKERRQ(ierr); 923 ierr = MatSetSizes(C_dense,A->rmap->n,matcoloring->ncolors,A->rmap->n,matcoloring->ncolors);CHKERRQ(ierr); 924 ierr = MatSetType(C_dense,MATSEQDENSE);CHKERRQ(ierr); 925 ierr = MatSeqDenseSetPreallocation(C_dense,NULL);CHKERRQ(ierr); 926 927 Bt_dense->assembled = PETSC_TRUE; 928 multtrans->ABt_den = C_dense; 929 930 #if defined(PETSC_USE_INFO) 931 { 932 Mat_SeqAIJ *c = (Mat_SeqAIJ*)(*C)->data; 933 ierr = PetscInfo5(*C,"Bt_dense: %D,%D; Cnz %D / (cm*ncolors %D) = %g\n",A->cmap->n,matcoloring->ncolors,c->nz,A->rmap->n*matcoloring->ncolors,(PetscReal)(c->nz)/(A->rmap->n*matcoloring->ncolors));CHKERRQ(ierr); 934 } 935 #endif 936 } 937 /* clean up */ 938 ierr = MatDestroy(&Bt);CHKERRQ(ierr); 939 ierr = MatRestoreSymbolicTranspose_SeqAIJ(B,&bti,&btj);CHKERRQ(ierr); 940 PetscFunctionReturn(0); 941 } 942 943 /* #define USE_ARRAY - for sparse dot product. Slower than !USE_ARRAY */ 944 #undef __FUNCT__ 945 #define __FUNCT__ "MatMatTransposeMultNumeric_SeqAIJ_SeqAIJ" 946 PetscErrorCode MatMatTransposeMultNumeric_SeqAIJ_SeqAIJ(Mat A,Mat B,Mat C) 947 { 948 PetscErrorCode ierr; 949 Mat_SeqAIJ *a =(Mat_SeqAIJ*)A->data,*b=(Mat_SeqAIJ*)B->data,*c=(Mat_SeqAIJ*)C->data; 950 PetscInt *ai =a->i,*aj=a->j,*bi=b->i,*bj=b->j,anzi,bnzj,nexta,nextb,*acol,*bcol,brow; 951 PetscInt cm =C->rmap->n,*ci=c->i,*cj=c->j,i,j,cnzi,*ccol; 952 PetscLogDouble flops=0.0; 953 MatScalar *aa =a->a,*aval,*ba=b->a,*bval,*ca,*cval; 954 Mat_MatMatTransMult *multtrans; 955 PetscContainer container; 956 #if defined(USE_ARRAY) 957 MatScalar *spdot; 958 #endif 959 960 PetscFunctionBegin; 961 /* clear old values in C */ 962 if (!c->a) { 963 ierr = PetscMalloc((ci[cm]+1)*sizeof(MatScalar),&ca);CHKERRQ(ierr); 964 c->a = ca; 965 c->free_a = PETSC_TRUE; 966 } else { 967 ca = c->a; 968 } 969 ierr = PetscMemzero(ca,ci[cm]*sizeof(MatScalar));CHKERRQ(ierr); 970 971 ierr = PetscObjectQuery((PetscObject)C,"Mat_MatMatTransMult",(PetscObject*)&container);CHKERRQ(ierr); 972 if (!container) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Container does not exit"); 973 ierr = PetscContainerGetPointer(container,(void**)&multtrans);CHKERRQ(ierr); 974 if (multtrans->usecoloring) { 975 MatTransposeColoring matcoloring = multtrans->matcoloring; 976 Mat Bt_dense; 977 PetscInt m,n; 978 Mat C_dense = multtrans->ABt_den; 979 980 Bt_dense = multtrans->Bt_den; 981 ierr = MatGetLocalSize(Bt_dense,&m,&n);CHKERRQ(ierr); 982 983 /* Get Bt_dense by Apply MatTransposeColoring to B */ 984 ierr = MatTransColoringApplySpToDen(matcoloring,B,Bt_dense);CHKERRQ(ierr); 985 986 /* C_dense = A*Bt_dense */ 987 ierr = MatMatMultNumeric_SeqAIJ_SeqDense(A,Bt_dense,C_dense);CHKERRQ(ierr); 988 989 /* Recover C from C_dense */ 990 ierr = MatTransColoringApplyDenToSp(matcoloring,C_dense,C);CHKERRQ(ierr); 991 PetscFunctionReturn(0); 992 } 993 994 #if defined(USE_ARRAY) 995 /* allocate an array for implementing sparse inner-product */ 996 ierr = PetscMalloc((A->cmap->n+1)*sizeof(MatScalar),&spdot);CHKERRQ(ierr); 997 ierr = PetscMemzero(spdot,(A->cmap->n+1)*sizeof(MatScalar));CHKERRQ(ierr); 998 #endif 999 1000 for (i=0; i<cm; i++) { 1001 anzi = ai[i+1] - ai[i]; 1002 acol = aj + ai[i]; 1003 aval = aa + ai[i]; 1004 cnzi = ci[i+1] - ci[i]; 1005 ccol = cj + ci[i]; 1006 cval = ca + ci[i]; 1007 for (j=0; j<cnzi; j++) { 1008 brow = ccol[j]; 1009 bnzj = bi[brow+1] - bi[brow]; 1010 bcol = bj + bi[brow]; 1011 bval = ba + bi[brow]; 1012 1013 /* perform sparse inner-product c(i,j)=A[i,:]*B[j,:]^T */ 1014 #if defined(USE_ARRAY) 1015 /* put ba to spdot array */ 1016 for (nextb=0; nextb<bnzj; nextb++) spdot[bcol[nextb]] = bval[nextb]; 1017 /* c(i,j)=A[i,:]*B[j,:]^T */ 1018 for (nexta=0; nexta<anzi; nexta++) { 1019 cval[j] += spdot[acol[nexta]]*aval[nexta]; 1020 } 1021 /* zero spdot array */ 1022 for (nextb=0; nextb<bnzj; nextb++) spdot[bcol[nextb]] = 0.0; 1023 #else 1024 nexta = 0; nextb = 0; 1025 while (nexta<anzi && nextb<bnzj) { 1026 while (nexta < anzi && acol[nexta] < bcol[nextb]) nexta++; 1027 if (nexta == anzi) break; 1028 while (nextb < bnzj && acol[nexta] > bcol[nextb]) nextb++; 1029 if (nextb == bnzj) break; 1030 if (acol[nexta] == bcol[nextb]) { 1031 cval[j] += aval[nexta]*bval[nextb]; 1032 nexta++; nextb++; 1033 flops += 2; 1034 } 1035 } 1036 #endif 1037 } 1038 } 1039 ierr = MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1040 ierr = MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1041 ierr = PetscLogFlops(flops);CHKERRQ(ierr); 1042 #if defined(USE_ARRAY) 1043 ierr = PetscFree(spdot); 1044 #endif 1045 PetscFunctionReturn(0); 1046 } 1047 1048 #undef __FUNCT__ 1049 #define __FUNCT__ "MatTransposeMatMult_SeqAIJ_SeqAIJ" 1050 PetscErrorCode MatTransposeMatMult_SeqAIJ_SeqAIJ(Mat A,Mat B,MatReuse scall,PetscReal fill,Mat *C) 1051 { 1052 PetscErrorCode ierr; 1053 1054 PetscFunctionBegin; 1055 if (scall == MAT_INITIAL_MATRIX) { 1056 ierr = MatTransposeMatMultSymbolic_SeqAIJ_SeqAIJ(A,B,fill,C);CHKERRQ(ierr); 1057 } 1058 ierr = MatTransposeMatMultNumeric_SeqAIJ_SeqAIJ(A,B,*C);CHKERRQ(ierr); 1059 PetscFunctionReturn(0); 1060 } 1061 1062 #undef __FUNCT__ 1063 #define __FUNCT__ "MatTransposeMatMultSymbolic_SeqAIJ_SeqAIJ" 1064 PetscErrorCode MatTransposeMatMultSymbolic_SeqAIJ_SeqAIJ(Mat A,Mat B,PetscReal fill,Mat *C) 1065 { 1066 PetscErrorCode ierr; 1067 Mat At; 1068 PetscInt *ati,*atj; 1069 1070 PetscFunctionBegin; 1071 /* create symbolic At */ 1072 ierr = MatGetSymbolicTranspose_SeqAIJ(A,&ati,&atj);CHKERRQ(ierr); 1073 ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,A->cmap->n,A->rmap->n,ati,atj,NULL,&At);CHKERRQ(ierr); 1074 1075 At->rmap->bs = A->cmap->bs; 1076 At->cmap->bs = B->cmap->bs; 1077 1078 /* get symbolic C=At*B */ 1079 ierr = MatMatMultSymbolic_SeqAIJ_SeqAIJ(At,B,fill,C);CHKERRQ(ierr); 1080 1081 /* clean up */ 1082 ierr = MatDestroy(&At);CHKERRQ(ierr); 1083 ierr = MatRestoreSymbolicTranspose_SeqAIJ(A,&ati,&atj);CHKERRQ(ierr); 1084 PetscFunctionReturn(0); 1085 } 1086 1087 #undef __FUNCT__ 1088 #define __FUNCT__ "MatTransposeMatMultNumeric_SeqAIJ_SeqAIJ" 1089 PetscErrorCode MatTransposeMatMultNumeric_SeqAIJ_SeqAIJ(Mat A,Mat B,Mat C) 1090 { 1091 PetscErrorCode ierr; 1092 Mat_SeqAIJ *a =(Mat_SeqAIJ*)A->data,*b=(Mat_SeqAIJ*)B->data,*c=(Mat_SeqAIJ*)C->data; 1093 PetscInt am =A->rmap->n,anzi,*ai=a->i,*aj=a->j,*bi=b->i,*bj,bnzi,nextb; 1094 PetscInt cm =C->rmap->n,*ci=c->i,*cj=c->j,crow,*cjj,i,j,k; 1095 PetscLogDouble flops=0.0; 1096 MatScalar *aa =a->a,*ba,*ca,*caj; 1097 1098 PetscFunctionBegin; 1099 if (!c->a) { 1100 ierr = PetscMalloc((ci[cm]+1)*sizeof(MatScalar),&ca);CHKERRQ(ierr); 1101 1102 c->a = ca; 1103 c->free_a = PETSC_TRUE; 1104 } else { 1105 ca = c->a; 1106 } 1107 /* clear old values in C */ 1108 ierr = PetscMemzero(ca,ci[cm]*sizeof(MatScalar));CHKERRQ(ierr); 1109 1110 /* compute A^T*B using outer product (A^T)[:,i]*B[i,:] */ 1111 for (i=0; i<am; i++) { 1112 bj = b->j + bi[i]; 1113 ba = b->a + bi[i]; 1114 bnzi = bi[i+1] - bi[i]; 1115 anzi = ai[i+1] - ai[i]; 1116 for (j=0; j<anzi; j++) { 1117 nextb = 0; 1118 crow = *aj++; 1119 cjj = cj + ci[crow]; 1120 caj = ca + ci[crow]; 1121 /* perform sparse axpy operation. Note cjj includes bj. */ 1122 for (k=0; nextb<bnzi; k++) { 1123 if (cjj[k] == *(bj+nextb)) { /* ccol == bcol */ 1124 caj[k] += (*aa)*(*(ba+nextb)); 1125 nextb++; 1126 } 1127 } 1128 flops += 2*bnzi; 1129 aa++; 1130 } 1131 } 1132 1133 /* Assemble the final matrix and clean up */ 1134 ierr = MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1135 ierr = MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1136 ierr = PetscLogFlops(flops);CHKERRQ(ierr); 1137 PetscFunctionReturn(0); 1138 } 1139 1140 #undef __FUNCT__ 1141 #define __FUNCT__ "MatMatMult_SeqAIJ_SeqDense" 1142 PetscErrorCode MatMatMult_SeqAIJ_SeqDense(Mat A,Mat B,MatReuse scall,PetscReal fill,Mat *C) 1143 { 1144 PetscErrorCode ierr; 1145 1146 PetscFunctionBegin; 1147 if (scall == MAT_INITIAL_MATRIX) { 1148 ierr = PetscLogEventBegin(MAT_MatMultSymbolic,A,B,0,0);CHKERRQ(ierr); 1149 ierr = MatMatMultSymbolic_SeqAIJ_SeqDense(A,B,fill,C);CHKERRQ(ierr); 1150 ierr = PetscLogEventEnd(MAT_MatMultSymbolic,A,B,0,0);CHKERRQ(ierr); 1151 } 1152 ierr = PetscLogEventBegin(MAT_MatMultNumeric,A,B,0,0);CHKERRQ(ierr); 1153 ierr = MatMatMultNumeric_SeqAIJ_SeqDense(A,B,*C);CHKERRQ(ierr); 1154 ierr = PetscLogEventEnd(MAT_MatMultNumeric,A,B,0,0);CHKERRQ(ierr); 1155 PetscFunctionReturn(0); 1156 } 1157 1158 #undef __FUNCT__ 1159 #define __FUNCT__ "MatMatMultSymbolic_SeqAIJ_SeqDense" 1160 PetscErrorCode MatMatMultSymbolic_SeqAIJ_SeqDense(Mat A,Mat B,PetscReal fill,Mat *C) 1161 { 1162 PetscErrorCode ierr; 1163 1164 PetscFunctionBegin; 1165 ierr = MatMatMultSymbolic_SeqDense_SeqDense(A,B,0.0,C);CHKERRQ(ierr); 1166 1167 (*C)->ops->matmultnumeric = MatMatMultNumeric_SeqAIJ_SeqDense; 1168 PetscFunctionReturn(0); 1169 } 1170 1171 #undef __FUNCT__ 1172 #define __FUNCT__ "MatMatMultNumeric_SeqAIJ_SeqDense" 1173 PetscErrorCode MatMatMultNumeric_SeqAIJ_SeqDense(Mat A,Mat B,Mat C) 1174 { 1175 Mat_SeqAIJ *a=(Mat_SeqAIJ*)A->data; 1176 PetscErrorCode ierr; 1177 PetscScalar *b,*c,r1,r2,r3,r4,*b1,*b2,*b3,*b4; 1178 MatScalar *aa; 1179 PetscInt cm = C->rmap->n, cn=B->cmap->n, bm=B->rmap->n, col, i,j,n,*aj, am = A->rmap->n; 1180 PetscInt am2 = 2*am, am3 = 3*am, bm4 = 4*bm,colam; 1181 1182 PetscFunctionBegin; 1183 if (!cm || !cn) PetscFunctionReturn(0); 1184 if (bm != A->cmap->n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Number columns in A %D not equal rows in B %D\n",A->cmap->n,bm); 1185 if (A->rmap->n != C->rmap->n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Number rows in C %D not equal rows in A %D\n",C->rmap->n,A->rmap->n); 1186 if (B->cmap->n != C->cmap->n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Number columns in B %D not equal columns in C %D\n",B->cmap->n,C->cmap->n); 1187 ierr = MatDenseGetArray(B,&b);CHKERRQ(ierr); 1188 ierr = MatDenseGetArray(C,&c);CHKERRQ(ierr); 1189 b1 = b; b2 = b1 + bm; b3 = b2 + bm; b4 = b3 + bm; 1190 for (col=0; col<cn-4; col += 4) { /* over columns of C */ 1191 colam = col*am; 1192 for (i=0; i<am; i++) { /* over rows of C in those columns */ 1193 r1 = r2 = r3 = r4 = 0.0; 1194 n = a->i[i+1] - a->i[i]; 1195 aj = a->j + a->i[i]; 1196 aa = a->a + a->i[i]; 1197 for (j=0; j<n; j++) { 1198 r1 += (*aa)*b1[*aj]; 1199 r2 += (*aa)*b2[*aj]; 1200 r3 += (*aa)*b3[*aj]; 1201 r4 += (*aa++)*b4[*aj++]; 1202 } 1203 c[colam + i] = r1; 1204 c[colam + am + i] = r2; 1205 c[colam + am2 + i] = r3; 1206 c[colam + am3 + i] = r4; 1207 } 1208 b1 += bm4; 1209 b2 += bm4; 1210 b3 += bm4; 1211 b4 += bm4; 1212 } 1213 for (; col<cn; col++) { /* over extra columns of C */ 1214 for (i=0; i<am; i++) { /* over rows of C in those columns */ 1215 r1 = 0.0; 1216 n = a->i[i+1] - a->i[i]; 1217 aj = a->j + a->i[i]; 1218 aa = a->a + a->i[i]; 1219 1220 for (j=0; j<n; j++) { 1221 r1 += (*aa++)*b1[*aj++]; 1222 } 1223 c[col*am + i] = r1; 1224 } 1225 b1 += bm; 1226 } 1227 ierr = PetscLogFlops(cn*(2.0*a->nz));CHKERRQ(ierr); 1228 ierr = MatDenseRestoreArray(B,&b);CHKERRQ(ierr); 1229 ierr = MatDenseRestoreArray(C,&c);CHKERRQ(ierr); 1230 ierr = MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1231 ierr = MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1232 PetscFunctionReturn(0); 1233 } 1234 1235 /* 1236 Note very similar to MatMult_SeqAIJ(), should generate both codes from same base 1237 */ 1238 #undef __FUNCT__ 1239 #define __FUNCT__ "MatMatMultNumericAdd_SeqAIJ_SeqDense" 1240 PetscErrorCode MatMatMultNumericAdd_SeqAIJ_SeqDense(Mat A,Mat B,Mat C) 1241 { 1242 Mat_SeqAIJ *a=(Mat_SeqAIJ*)A->data; 1243 PetscErrorCode ierr; 1244 PetscScalar *b,*c,r1,r2,r3,r4,*b1,*b2,*b3,*b4; 1245 MatScalar *aa; 1246 PetscInt cm = C->rmap->n, cn=B->cmap->n, bm=B->rmap->n, col, i,j,n,*aj, am = A->rmap->n,*ii,arm; 1247 PetscInt am2 = 2*am, am3 = 3*am, bm4 = 4*bm,colam,*ridx; 1248 1249 PetscFunctionBegin; 1250 if (!cm || !cn) PetscFunctionReturn(0); 1251 ierr = MatDenseGetArray(B,&b);CHKERRQ(ierr); 1252 ierr = MatDenseGetArray(C,&c);CHKERRQ(ierr); 1253 b1 = b; b2 = b1 + bm; b3 = b2 + bm; b4 = b3 + bm; 1254 1255 if (a->compressedrow.use) { /* use compressed row format */ 1256 for (col=0; col<cn-4; col += 4) { /* over columns of C */ 1257 colam = col*am; 1258 arm = a->compressedrow.nrows; 1259 ii = a->compressedrow.i; 1260 ridx = a->compressedrow.rindex; 1261 for (i=0; i<arm; i++) { /* over rows of C in those columns */ 1262 r1 = r2 = r3 = r4 = 0.0; 1263 n = ii[i+1] - ii[i]; 1264 aj = a->j + ii[i]; 1265 aa = a->a + ii[i]; 1266 for (j=0; j<n; j++) { 1267 r1 += (*aa)*b1[*aj]; 1268 r2 += (*aa)*b2[*aj]; 1269 r3 += (*aa)*b3[*aj]; 1270 r4 += (*aa++)*b4[*aj++]; 1271 } 1272 c[colam + ridx[i]] += r1; 1273 c[colam + am + ridx[i]] += r2; 1274 c[colam + am2 + ridx[i]] += r3; 1275 c[colam + am3 + ridx[i]] += r4; 1276 } 1277 b1 += bm4; 1278 b2 += bm4; 1279 b3 += bm4; 1280 b4 += bm4; 1281 } 1282 for (; col<cn; col++) { /* over extra columns of C */ 1283 colam = col*am; 1284 arm = a->compressedrow.nrows; 1285 ii = a->compressedrow.i; 1286 ridx = a->compressedrow.rindex; 1287 for (i=0; i<arm; i++) { /* over rows of C in those columns */ 1288 r1 = 0.0; 1289 n = ii[i+1] - ii[i]; 1290 aj = a->j + ii[i]; 1291 aa = a->a + ii[i]; 1292 1293 for (j=0; j<n; j++) { 1294 r1 += (*aa++)*b1[*aj++]; 1295 } 1296 c[colam + ridx[i]] += r1; 1297 } 1298 b1 += bm; 1299 } 1300 } else { 1301 for (col=0; col<cn-4; col += 4) { /* over columns of C */ 1302 colam = col*am; 1303 for (i=0; i<am; i++) { /* over rows of C in those columns */ 1304 r1 = r2 = r3 = r4 = 0.0; 1305 n = a->i[i+1] - a->i[i]; 1306 aj = a->j + a->i[i]; 1307 aa = a->a + a->i[i]; 1308 for (j=0; j<n; j++) { 1309 r1 += (*aa)*b1[*aj]; 1310 r2 += (*aa)*b2[*aj]; 1311 r3 += (*aa)*b3[*aj]; 1312 r4 += (*aa++)*b4[*aj++]; 1313 } 1314 c[colam + i] += r1; 1315 c[colam + am + i] += r2; 1316 c[colam + am2 + i] += r3; 1317 c[colam + am3 + i] += r4; 1318 } 1319 b1 += bm4; 1320 b2 += bm4; 1321 b3 += bm4; 1322 b4 += bm4; 1323 } 1324 for (; col<cn; col++) { /* over extra columns of C */ 1325 colam = col*am; 1326 for (i=0; i<am; i++) { /* over rows of C in those columns */ 1327 r1 = 0.0; 1328 n = a->i[i+1] - a->i[i]; 1329 aj = a->j + a->i[i]; 1330 aa = a->a + a->i[i]; 1331 1332 for (j=0; j<n; j++) { 1333 r1 += (*aa++)*b1[*aj++]; 1334 } 1335 c[colam + i] += r1; 1336 } 1337 b1 += bm; 1338 } 1339 } 1340 ierr = PetscLogFlops(cn*2.0*a->nz);CHKERRQ(ierr); 1341 ierr = MatDenseRestoreArray(B,&b);CHKERRQ(ierr); 1342 ierr = MatDenseRestoreArray(C,&c);CHKERRQ(ierr); 1343 PetscFunctionReturn(0); 1344 } 1345 1346 #undef __FUNCT__ 1347 #define __FUNCT__ "MatTransColoringApplySpToDen_SeqAIJ" 1348 PetscErrorCode MatTransColoringApplySpToDen_SeqAIJ(MatTransposeColoring coloring,Mat B,Mat Btdense) 1349 { 1350 PetscErrorCode ierr; 1351 Mat_SeqAIJ *b = (Mat_SeqAIJ*)B->data; 1352 Mat_SeqDense *btdense = (Mat_SeqDense*)Btdense->data; 1353 PetscInt *bi = b->i,*bj=b->j; 1354 PetscInt m = Btdense->rmap->n,n=Btdense->cmap->n,j,k,l,col,anz,*btcol,brow,ncolumns; 1355 MatScalar *btval,*btval_den,*ba=b->a; 1356 PetscInt *columns=coloring->columns,*colorforcol=coloring->colorforcol,ncolors=coloring->ncolors; 1357 1358 PetscFunctionBegin; 1359 btval_den=btdense->v; 1360 ierr = PetscMemzero(btval_den,(m*n)*sizeof(MatScalar));CHKERRQ(ierr); 1361 for (k=0; k<ncolors; k++) { 1362 ncolumns = coloring->ncolumns[k]; 1363 for (l=0; l<ncolumns; l++) { /* insert a row of B to a column of Btdense */ 1364 col = *(columns + colorforcol[k] + l); 1365 btcol = bj + bi[col]; 1366 btval = ba + bi[col]; 1367 anz = bi[col+1] - bi[col]; 1368 for (j=0; j<anz; j++) { 1369 brow = btcol[j]; 1370 btval_den[brow] = btval[j]; 1371 } 1372 } 1373 btval_den += m; 1374 } 1375 PetscFunctionReturn(0); 1376 } 1377 1378 #undef __FUNCT__ 1379 #define __FUNCT__ "MatTransColoringApplyDenToSp_SeqAIJ" 1380 PetscErrorCode MatTransColoringApplyDenToSp_SeqAIJ(MatTransposeColoring matcoloring,Mat Cden,Mat Csp) 1381 { 1382 PetscErrorCode ierr; 1383 Mat_SeqAIJ *csp = (Mat_SeqAIJ*)Csp->data; 1384 PetscInt k,l,*row,*idx,m,ncolors=matcoloring->ncolors,nrows; 1385 PetscScalar *ca_den,*cp_den,*ca=csp->a; 1386 PetscInt *rows=matcoloring->rows,*spidx=matcoloring->columnsforspidx,*colorforrow=matcoloring->colorforrow; 1387 1388 PetscFunctionBegin; 1389 ierr = MatGetLocalSize(Csp,&m,NULL);CHKERRQ(ierr); 1390 ierr = MatDenseGetArray(Cden,&ca_den);CHKERRQ(ierr); 1391 cp_den = ca_den; 1392 for (k=0; k<ncolors; k++) { 1393 nrows = matcoloring->nrows[k]; 1394 row = rows + colorforrow[k]; 1395 idx = spidx + colorforrow[k]; 1396 for (l=0; l<nrows; l++) { 1397 ca[idx[l]] = cp_den[row[l]]; 1398 } 1399 cp_den += m; 1400 } 1401 ierr = MatDenseRestoreArray(Cden,&ca_den);CHKERRQ(ierr); 1402 PetscFunctionReturn(0); 1403 } 1404 1405 /* 1406 MatGetColumnIJ_SeqAIJ_Color() and MatRestoreColumnIJ_SeqAIJ_Color() are customized from 1407 MatGetColumnIJ_SeqAIJ() and MatRestoreColumnIJ_SeqAIJ() by adding an output 1408 spidx[], index of a->j, to be used for setting 'columnsforspidx' in MatTransposeColoringCreate_SeqAIJ(). 1409 */ 1410 #undef __FUNCT__ 1411 #define __FUNCT__ "MatGetColumnIJ_SeqAIJ_Color" 1412 PetscErrorCode MatGetColumnIJ_SeqAIJ_Color(Mat A,PetscInt oshift,PetscBool symmetric,PetscBool inodecompressed,PetscInt *nn,const PetscInt *ia[],const PetscInt *ja[],PetscInt *spidx[],PetscBool *done) 1413 { 1414 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 1415 PetscErrorCode ierr; 1416 PetscInt i,*collengths,*cia,*cja,n = A->cmap->n,m = A->rmap->n; 1417 PetscInt nz = a->i[m],row,*jj,mr,col; 1418 PetscInt *cspidx; 1419 1420 PetscFunctionBegin; 1421 *nn = n; 1422 if (!ia) PetscFunctionReturn(0); 1423 if (symmetric) { 1424 SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"MatGetColumnIJ_SeqAIJ_Color() not supported for the case symmetric"); 1425 ierr = MatToSymmetricIJ_SeqAIJ(A->rmap->n,a->i,a->j,0,oshift,(PetscInt**)ia,(PetscInt**)ja);CHKERRQ(ierr); 1426 } else { 1427 ierr = PetscMalloc((n+1)*sizeof(PetscInt),&collengths);CHKERRQ(ierr); 1428 ierr = PetscMemzero(collengths,n*sizeof(PetscInt));CHKERRQ(ierr); 1429 ierr = PetscMalloc((n+1)*sizeof(PetscInt),&cia);CHKERRQ(ierr); 1430 ierr = PetscMalloc((nz+1)*sizeof(PetscInt),&cja);CHKERRQ(ierr); 1431 ierr = PetscMalloc((nz+1)*sizeof(PetscInt),&cspidx);CHKERRQ(ierr); 1432 jj = a->j; 1433 for (i=0; i<nz; i++) { 1434 collengths[jj[i]]++; 1435 } 1436 cia[0] = oshift; 1437 for (i=0; i<n; i++) { 1438 cia[i+1] = cia[i] + collengths[i]; 1439 } 1440 ierr = PetscMemzero(collengths,n*sizeof(PetscInt));CHKERRQ(ierr); 1441 jj = a->j; 1442 for (row=0; row<m; row++) { 1443 mr = a->i[row+1] - a->i[row]; 1444 for (i=0; i<mr; i++) { 1445 col = *jj++; 1446 1447 cspidx[cia[col] + collengths[col] - oshift] = a->i[row] + i; /* index of a->j */ 1448 cja[cia[col] + collengths[col]++ - oshift] = row + oshift; 1449 } 1450 } 1451 ierr = PetscFree(collengths);CHKERRQ(ierr); 1452 *ia = cia; *ja = cja; 1453 *spidx = cspidx; 1454 } 1455 PetscFunctionReturn(0); 1456 } 1457 1458 #undef __FUNCT__ 1459 #define __FUNCT__ "MatRestoreColumnIJ_SeqAIJ_Color" 1460 PetscErrorCode MatRestoreColumnIJ_SeqAIJ_Color(Mat A,PetscInt oshift,PetscBool symmetric,PetscBool inodecompressed,PetscInt *n,const PetscInt *ia[],const PetscInt *ja[],PetscInt *spidx[],PetscBool *done) 1461 { 1462 PetscErrorCode ierr; 1463 1464 PetscFunctionBegin; 1465 if (!ia) PetscFunctionReturn(0); 1466 1467 ierr = PetscFree(*ia);CHKERRQ(ierr); 1468 ierr = PetscFree(*ja);CHKERRQ(ierr); 1469 ierr = PetscFree(*spidx);CHKERRQ(ierr); 1470 PetscFunctionReturn(0); 1471 } 1472 1473 #undef __FUNCT__ 1474 #define __FUNCT__ "MatTransposeColoringCreate_SeqAIJ" 1475 PetscErrorCode MatTransposeColoringCreate_SeqAIJ(Mat mat,ISColoring iscoloring,MatTransposeColoring c) 1476 { 1477 PetscErrorCode ierr; 1478 PetscInt i,n,nrows,N,j,k,m,ncols,col,cm; 1479 const PetscInt *is,*ci,*cj,*row_idx; 1480 PetscInt nis = iscoloring->n,*rowhit,bs = 1; 1481 IS *isa; 1482 PetscBool flg1,flg2; 1483 Mat_SeqAIJ *csp = (Mat_SeqAIJ*)mat->data; 1484 PetscInt *colorforrow,*rows,*rows_i,*columnsforspidx,*columnsforspidx_i,*idxhit,*spidx; 1485 PetscInt *colorforcol,*columns,*columns_i; 1486 1487 PetscFunctionBegin; 1488 ierr = ISColoringGetIS(iscoloring,PETSC_IGNORE,&isa);CHKERRQ(ierr); 1489 1490 /* this is ugly way to get blocksize but cannot call MatGetBlockSize() because AIJ can have bs > 1 */ 1491 ierr = PetscObjectTypeCompare((PetscObject)mat,MATSEQBAIJ,&flg1);CHKERRQ(ierr); 1492 ierr = PetscObjectTypeCompare((PetscObject)mat,MATMPIBAIJ,&flg2);CHKERRQ(ierr); 1493 if (flg1 || flg2) { 1494 ierr = MatGetBlockSize(mat,&bs);CHKERRQ(ierr); 1495 } 1496 1497 N = mat->cmap->N/bs; 1498 c->M = mat->rmap->N/bs; /* set total rows, columns and local rows */ 1499 c->N = mat->cmap->N/bs; 1500 c->m = mat->rmap->N/bs; 1501 c->rstart = 0; 1502 1503 c->ncolors = nis; 1504 ierr = PetscMalloc(nis*sizeof(PetscInt),&c->ncolumns);CHKERRQ(ierr); 1505 ierr = PetscMalloc(nis*sizeof(PetscInt),&c->nrows);CHKERRQ(ierr); 1506 ierr = PetscMalloc2(csp->nz+1,PetscInt,&rows,csp->nz+1,PetscInt,&columnsforspidx);CHKERRQ(ierr); 1507 ierr = PetscMalloc((nis+1)*sizeof(PetscInt),&colorforrow);CHKERRQ(ierr); 1508 1509 colorforrow[0] = 0; 1510 rows_i = rows; 1511 columnsforspidx_i = columnsforspidx; 1512 1513 ierr = PetscMalloc((nis+1)*sizeof(PetscInt),&colorforcol);CHKERRQ(ierr); 1514 ierr = PetscMalloc((N+1)*sizeof(PetscInt),&columns);CHKERRQ(ierr); 1515 1516 colorforcol[0] = 0; 1517 columns_i = columns; 1518 1519 ierr = MatGetColumnIJ_SeqAIJ_Color(mat,0,PETSC_FALSE,PETSC_FALSE,&ncols,&ci,&cj,&spidx,NULL);CHKERRQ(ierr); /* column-wise storage of mat */ 1520 1521 cm = c->m; 1522 ierr = PetscMalloc((cm+1)*sizeof(PetscInt),&rowhit);CHKERRQ(ierr); 1523 ierr = PetscMalloc((cm+1)*sizeof(PetscInt),&idxhit);CHKERRQ(ierr); 1524 for (i=0; i<nis; i++) { 1525 ierr = ISGetLocalSize(isa[i],&n);CHKERRQ(ierr); 1526 ierr = ISGetIndices(isa[i],&is);CHKERRQ(ierr); 1527 1528 c->ncolumns[i] = n; 1529 if (n) { 1530 ierr = PetscMemcpy(columns_i,is,n*sizeof(PetscInt));CHKERRQ(ierr); 1531 } 1532 colorforcol[i+1] = colorforcol[i] + n; 1533 columns_i += n; 1534 1535 /* fast, crude version requires O(N*N) work */ 1536 ierr = PetscMemzero(rowhit,cm*sizeof(PetscInt));CHKERRQ(ierr); 1537 1538 /* loop over columns*/ 1539 for (j=0; j<n; j++) { 1540 col = is[j]; 1541 row_idx = cj + ci[col]; 1542 m = ci[col+1] - ci[col]; 1543 /* loop over columns marking them in rowhit */ 1544 for (k=0; k<m; k++) { 1545 idxhit[*row_idx] = spidx[ci[col] + k]; 1546 rowhit[*row_idx++] = col + 1; 1547 } 1548 } 1549 /* count the number of hits */ 1550 nrows = 0; 1551 for (j=0; j<cm; j++) { 1552 if (rowhit[j]) nrows++; 1553 } 1554 c->nrows[i] = nrows; 1555 colorforrow[i+1] = colorforrow[i] + nrows; 1556 1557 nrows = 0; 1558 for (j=0; j<cm; j++) { 1559 if (rowhit[j]) { 1560 rows_i[nrows] = j; 1561 columnsforspidx_i[nrows] = idxhit[j]; 1562 nrows++; 1563 } 1564 } 1565 ierr = ISRestoreIndices(isa[i],&is);CHKERRQ(ierr); 1566 rows_i += nrows; columnsforspidx_i += nrows; 1567 } 1568 ierr = MatRestoreColumnIJ_SeqAIJ_Color(mat,0,PETSC_FALSE,PETSC_FALSE,&ncols,&ci,&cj,&spidx,NULL);CHKERRQ(ierr); 1569 ierr = PetscFree(rowhit);CHKERRQ(ierr); 1570 ierr = ISColoringRestoreIS(iscoloring,&isa);CHKERRQ(ierr); 1571 #if defined(PETSC_USE_DEBUG) 1572 if (csp->nz != colorforrow[nis]) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"csp->nz %d != colorforrow[nis] %d",csp->nz,colorforrow[nis]); 1573 #endif 1574 1575 c->colorforrow = colorforrow; 1576 c->rows = rows; 1577 c->columnsforspidx = columnsforspidx; 1578 c->colorforcol = colorforcol; 1579 c->columns = columns; 1580 1581 ierr = PetscFree(idxhit);CHKERRQ(ierr); 1582 PetscFunctionReturn(0); 1583 } 1584