1 2 /*************************************xxt.c************************************ 3 Module Name: xxt 4 Module Info: 5 6 author: Henry M. Tufo III 7 e-mail: hmt@asci.uchicago.edu 8 contact: 9 +--------------------------------+--------------------------------+ 10 |MCS Division - Building 221 |Department of Computer Science | 11 |Argonne National Laboratory |Ryerson 152 | 12 |9700 S. Cass Avenue |The University of Chicago | 13 |Argonne, IL 60439 |Chicago, IL 60637 | 14 |(630) 252-5354/5986 ph/fx |(773) 702-6019/8487 ph/fx | 15 +--------------------------------+--------------------------------+ 16 17 Last Modification: 3.20.01 18 **************************************xxt.c***********************************/ 19 #include <../src/ksp/pc/impls/tfs/tfs.h> 20 21 #define LEFT -1 22 #define RIGHT 1 23 #define BOTH 0 24 25 typedef struct xxt_solver_info { 26 PetscInt n, m, n_global, m_global; 27 PetscInt nnz, max_nnz, msg_buf_sz; 28 PetscInt *nsep, *lnsep, *fo, nfo, *stages; 29 PetscInt *col_sz, *col_indices; 30 PetscScalar **col_vals, *x, *solve_uu, *solve_w; 31 PetscInt nsolves; 32 PetscScalar tot_solve_time; 33 } xxt_info; 34 35 typedef struct matvec_info { 36 PetscInt n, m, n_global, m_global; 37 PetscInt *local2global; 38 PCTFS_gs_ADT PCTFS_gs_handle; 39 PetscErrorCode (*matvec)(struct matvec_info*,PetscScalar*,PetscScalar*); 40 void *grid_data; 41 } mv_info; 42 43 struct xxt_CDT { 44 PetscInt id; 45 PetscInt ns; 46 PetscInt level; 47 xxt_info *info; 48 mv_info *mvi; 49 }; 50 51 static PetscInt n_xxt =0; 52 static PetscInt n_xxt_handles=0; 53 54 /* prototypes */ 55 static PetscErrorCode do_xxt_solve(xxt_ADT xxt_handle, PetscScalar *rhs); 56 static PetscErrorCode check_handle(xxt_ADT xxt_handle); 57 static PetscErrorCode det_separators(xxt_ADT xxt_handle); 58 static PetscErrorCode do_matvec(mv_info *A, PetscScalar *v, PetscScalar *u); 59 static PetscErrorCode xxt_generate(xxt_ADT xxt_handle); 60 static PetscErrorCode do_xxt_factor(xxt_ADT xxt_handle); 61 static mv_info *set_mvi(PetscInt *local2global, PetscInt n, PetscInt m, PetscErrorCode (*matvec)(mv_info*,PetscScalar*,PetscScalar*), void *grid_data); 62 63 /**************************************xxt.c***********************************/ 64 xxt_ADT XXT_new(void) 65 { 66 xxt_ADT xxt_handle; 67 68 /* rolling count on n_xxt ... pot. problem here */ 69 n_xxt_handles++; 70 xxt_handle = (xxt_ADT)malloc(sizeof(struct xxt_CDT)); 71 xxt_handle->id = ++n_xxt; 72 xxt_handle->info = NULL; xxt_handle->mvi = NULL; 73 74 return(xxt_handle); 75 } 76 77 /**************************************xxt.c***********************************/ 78 PetscErrorCode XXT_factor(xxt_ADT xxt_handle, /* prev. allocated xxt handle */ 79 PetscInt *local2global, /* global column mapping */ 80 PetscInt n, /* local num rows */ 81 PetscInt m, /* local num cols */ 82 PetscErrorCode (*matvec)(void*,PetscScalar*,PetscScalar*), /* b_loc=A_local.x_loc */ 83 void *grid_data) /* grid data for matvec */ 84 { 85 PCTFS_comm_init(); 86 check_handle(xxt_handle); 87 88 /* only 2^k for now and all nodes participating */ 89 if ((1<<(xxt_handle->level=PCTFS_i_log2_num_nodes))!=PCTFS_num_nodes) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"only 2^k for now and MPI_COMM_WORLD!!! %D != %D\n",1<<PCTFS_i_log2_num_nodes,PCTFS_num_nodes); 90 91 /* space for X info */ 92 xxt_handle->info = (xxt_info*)malloc(sizeof(xxt_info)); 93 94 /* set up matvec handles */ 95 xxt_handle->mvi = set_mvi(local2global, n, m, (PetscErrorCode (*)(mv_info*,PetscScalar*,PetscScalar*))matvec, grid_data); 96 97 /* matrix is assumed to be of full rank */ 98 /* LATER we can reset to indicate rank def. */ 99 xxt_handle->ns=0; 100 101 /* determine separators and generate firing order - NB xxt info set here */ 102 det_separators(xxt_handle); 103 104 return(do_xxt_factor(xxt_handle)); 105 } 106 107 /**************************************xxt.c***********************************/ 108 PetscErrorCode XXT_solve(xxt_ADT xxt_handle, PetscScalar *x, PetscScalar *b) 109 { 110 111 PCTFS_comm_init(); 112 check_handle(xxt_handle); 113 114 /* need to copy b into x? */ 115 if (b) PCTFS_rvec_copy(x,b,xxt_handle->mvi->n); 116 return do_xxt_solve(xxt_handle,x); 117 } 118 119 /**************************************xxt.c***********************************/ 120 PetscInt XXT_free(xxt_ADT xxt_handle) 121 { 122 123 PCTFS_comm_init(); 124 check_handle(xxt_handle); 125 n_xxt_handles--; 126 127 free(xxt_handle->info->nsep); 128 free(xxt_handle->info->lnsep); 129 free(xxt_handle->info->fo); 130 free(xxt_handle->info->stages); 131 free(xxt_handle->info->solve_uu); 132 free(xxt_handle->info->solve_w); 133 free(xxt_handle->info->x); 134 free(xxt_handle->info->col_vals); 135 free(xxt_handle->info->col_sz); 136 free(xxt_handle->info->col_indices); 137 free(xxt_handle->info); 138 free(xxt_handle->mvi->local2global); 139 PCTFS_gs_free(xxt_handle->mvi->PCTFS_gs_handle); 140 free(xxt_handle->mvi); 141 free(xxt_handle); 142 143 /* if the check fails we nuke */ 144 /* if NULL pointer passed to free we nuke */ 145 /* if the calls to free fail that's not my problem */ 146 return(0); 147 } 148 149 /**************************************xxt.c***********************************/ 150 PetscErrorCode XXT_stats(xxt_ADT xxt_handle) 151 { 152 PetscInt op[] = {NON_UNIFORM,GL_MIN,GL_MAX,GL_ADD,GL_MIN,GL_MAX,GL_ADD,GL_MIN,GL_MAX,GL_ADD}; 153 PetscInt fop[] = {NON_UNIFORM,GL_MIN,GL_MAX,GL_ADD}; 154 PetscInt vals[9], work[9]; 155 PetscScalar fvals[3], fwork[3]; 156 PetscErrorCode ierr; 157 158 PCTFS_comm_init(); 159 check_handle(xxt_handle); 160 161 /* if factorization not done there are no stats */ 162 if (!xxt_handle->info||!xxt_handle->mvi) { 163 if (!PCTFS_my_id) { ierr = PetscPrintf(PETSC_COMM_WORLD,"XXT_stats() :: no stats available!\n");CHKERRQ(ierr); } 164 return 1; 165 } 166 167 vals[0]=vals[1]=vals[2]=xxt_handle->info->nnz; 168 vals[3]=vals[4]=vals[5]=xxt_handle->mvi->n; 169 vals[6]=vals[7]=vals[8]=xxt_handle->info->msg_buf_sz; 170 PCTFS_giop(vals,work,sizeof(op)/sizeof(op[0])-1,op); 171 172 fvals[0]=fvals[1]=fvals[2] =xxt_handle->info->tot_solve_time/xxt_handle->info->nsolves++; 173 PCTFS_grop(fvals,fwork,sizeof(fop)/sizeof(fop[0])-1,fop); 174 175 if (!PCTFS_my_id) { 176 ierr = PetscPrintf(PETSC_COMM_WORLD,"%D :: min xxt_nnz=%D\n",PCTFS_my_id,vals[0]);CHKERRQ(ierr); 177 ierr = PetscPrintf(PETSC_COMM_WORLD,"%D :: max xxt_nnz=%D\n",PCTFS_my_id,vals[1]);CHKERRQ(ierr); 178 ierr = PetscPrintf(PETSC_COMM_WORLD,"%D :: avg xxt_nnz=%g\n",PCTFS_my_id,1.0*vals[2]/PCTFS_num_nodes);CHKERRQ(ierr); 179 ierr = PetscPrintf(PETSC_COMM_WORLD,"%D :: tot xxt_nnz=%D\n",PCTFS_my_id,vals[2]);CHKERRQ(ierr); 180 ierr = PetscPrintf(PETSC_COMM_WORLD,"%D :: xxt C(2d) =%g\n",PCTFS_my_id,vals[2]/(PetscPowReal(1.0*vals[5],1.5)));CHKERRQ(ierr); 181 ierr = PetscPrintf(PETSC_COMM_WORLD,"%D :: xxt C(3d) =%g\n",PCTFS_my_id,vals[2]/(PetscPowReal(1.0*vals[5],1.6667)));CHKERRQ(ierr); 182 ierr = PetscPrintf(PETSC_COMM_WORLD,"%D :: min xxt_n =%D\n",PCTFS_my_id,vals[3]);CHKERRQ(ierr); 183 ierr = PetscPrintf(PETSC_COMM_WORLD,"%D :: max xxt_n =%D\n",PCTFS_my_id,vals[4]);CHKERRQ(ierr); 184 ierr = PetscPrintf(PETSC_COMM_WORLD,"%D :: avg xxt_n =%g\n",PCTFS_my_id,1.0*vals[5]/PCTFS_num_nodes);CHKERRQ(ierr); 185 ierr = PetscPrintf(PETSC_COMM_WORLD,"%D :: tot xxt_n =%D\n",PCTFS_my_id,vals[5]);CHKERRQ(ierr); 186 ierr = PetscPrintf(PETSC_COMM_WORLD,"%D :: min xxt_buf=%D\n",PCTFS_my_id,vals[6]);CHKERRQ(ierr); 187 ierr = PetscPrintf(PETSC_COMM_WORLD,"%D :: max xxt_buf=%D\n",PCTFS_my_id,vals[7]);CHKERRQ(ierr); 188 ierr = PetscPrintf(PETSC_COMM_WORLD,"%D :: avg xxt_buf=%g\n",PCTFS_my_id,1.0*vals[8]/PCTFS_num_nodes);CHKERRQ(ierr); 189 ierr = PetscPrintf(PETSC_COMM_WORLD,"%D :: min xxt_slv=%g\n",PCTFS_my_id,fvals[0]);CHKERRQ(ierr); 190 ierr = PetscPrintf(PETSC_COMM_WORLD,"%D :: max xxt_slv=%g\n",PCTFS_my_id,fvals[1]);CHKERRQ(ierr); 191 ierr = PetscPrintf(PETSC_COMM_WORLD,"%D :: avg xxt_slv=%g\n",PCTFS_my_id,fvals[2]/PCTFS_num_nodes);CHKERRQ(ierr); 192 } 193 194 return(0); 195 } 196 197 /*************************************xxt.c************************************ 198 199 Description: get A_local, local portion of global coarse matrix which 200 is a row dist. nxm matrix w/ n<m. 201 o my_ml holds address of ML struct associated w/A_local and coarse grid 202 o local2global holds global number of column i (i=0,...,m-1) 203 o local2global holds global number of row i (i=0,...,n-1) 204 o mylocmatvec performs A_local . vec_local (note that gs is performed using 205 PCTFS_gs_init/gop). 206 207 mylocmatvec = my_ml->Amat[grid_tag].matvec->external; 208 mylocmatvec (void :: void *data, double *in, double *out) 209 **************************************xxt.c***********************************/ 210 static PetscErrorCode do_xxt_factor(xxt_ADT xxt_handle) 211 { 212 return xxt_generate(xxt_handle); 213 } 214 215 /**************************************xxt.c***********************************/ 216 static PetscErrorCode xxt_generate(xxt_ADT xxt_handle) 217 { 218 PetscInt i,j,k,idex; 219 PetscInt dim, col; 220 PetscScalar *u, *uu, *v, *z, *w, alpha, alpha_w; 221 PetscInt *segs; 222 PetscInt op[] = {GL_ADD,0}; 223 PetscInt off, len; 224 PetscScalar *x_ptr; 225 PetscInt *iptr, flag; 226 PetscInt start =0, end, work; 227 PetscInt op2[] = {GL_MIN,0}; 228 PCTFS_gs_ADT PCTFS_gs_handle; 229 PetscInt *nsep, *lnsep, *fo; 230 PetscInt a_n =xxt_handle->mvi->n; 231 PetscInt a_m =xxt_handle->mvi->m; 232 PetscInt *a_local2global=xxt_handle->mvi->local2global; 233 PetscInt level; 234 PetscInt xxt_nnz=0, xxt_max_nnz=0; 235 PetscInt n, m; 236 PetscInt *col_sz, *col_indices, *stages; 237 PetscScalar **col_vals, *x; 238 PetscInt n_global; 239 PetscInt xxt_zero_nnz =0; 240 PetscInt xxt_zero_nnz_0=0; 241 PetscBLASInt i1 = 1,dlen; 242 PetscScalar dm1 = -1.0; 243 PetscErrorCode ierr; 244 245 n = xxt_handle->mvi->n; 246 nsep = xxt_handle->info->nsep; 247 lnsep = xxt_handle->info->lnsep; 248 fo = xxt_handle->info->fo; 249 end = lnsep[0]; 250 level = xxt_handle->level; 251 PCTFS_gs_handle = xxt_handle->mvi->PCTFS_gs_handle; 252 253 /* is there a null space? */ 254 /* LATER add in ability to detect null space by checking alpha */ 255 for (i=0, j=0; i<=level; i++) j+=nsep[i]; 256 257 m = j-xxt_handle->ns; 258 if (m!=j) { 259 ierr = PetscPrintf(PETSC_COMM_WORLD,"xxt_generate() :: null space exists %D %D %D\n",m,j,xxt_handle->ns);CHKERRQ(ierr); 260 } 261 262 /* get and initialize storage for x local */ 263 /* note that x local is nxm and stored by columns */ 264 col_sz = (PetscInt*) malloc(m*sizeof(PetscInt)); 265 col_indices = (PetscInt*) malloc((2*m+1)*sizeof(PetscInt)); 266 col_vals = (PetscScalar**) malloc(m*sizeof(PetscScalar*)); 267 for (i=j=0; i<m; i++, j+=2) { 268 col_indices[j]=col_indices[j+1]=col_sz[i]=-1; 269 col_vals[i] = NULL; 270 } 271 col_indices[j]=-1; 272 273 /* size of separators for each sub-hc working from bottom of tree to top */ 274 /* this looks like nsep[]=segments */ 275 stages = (PetscInt*) malloc((level+1)*sizeof(PetscInt)); 276 segs = (PetscInt*) malloc((level+1)*sizeof(PetscInt)); 277 PCTFS_ivec_zero(stages,level+1); 278 PCTFS_ivec_copy(segs,nsep,level+1); 279 for (i=0; i<level; i++) segs[i+1] += segs[i]; 280 stages[0] = segs[0]; 281 282 /* temporary vectors */ 283 u = (PetscScalar*) malloc(n*sizeof(PetscScalar)); 284 z = (PetscScalar*) malloc(n*sizeof(PetscScalar)); 285 v = (PetscScalar*) malloc(a_m*sizeof(PetscScalar)); 286 uu = (PetscScalar*) malloc(m*sizeof(PetscScalar)); 287 w = (PetscScalar*) malloc(m*sizeof(PetscScalar)); 288 289 /* extra nnz due to replication of vertices across separators */ 290 for (i=1, j=0; i<=level; i++) j+=nsep[i]; 291 292 /* storage for sparse x values */ 293 n_global = xxt_handle->info->n_global; 294 xxt_max_nnz = (PetscInt)(2.5*PetscPowReal(1.0*n_global,1.6667) + j*n/2)/PCTFS_num_nodes; 295 x = (PetscScalar*) malloc(xxt_max_nnz*sizeof(PetscScalar)); 296 xxt_nnz = 0; 297 298 /* LATER - can embed next sep to fire in gs */ 299 /* time to make the donuts - generate X factor */ 300 for (dim=i=j=0; i<m; i++) { 301 /* time to move to the next level? */ 302 while (i==segs[dim]) { 303 if (dim==level) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"dim about to exceed level\n"); 304 stages[dim++]=i; 305 end +=lnsep[dim]; 306 } 307 stages[dim]=i; 308 309 /* which column are we firing? */ 310 /* i.e. set v_l */ 311 /* use new seps and do global min across hc to determine which one to fire */ 312 (start<end) ? (col=fo[start]) : (col=INT_MAX); 313 PCTFS_giop_hc(&col,&work,1,op2,dim); 314 315 /* shouldn't need this */ 316 if (col==INT_MAX) { 317 ierr = PetscInfo(0,"hey ... col==INT_MAX??\n");CHKERRQ(ierr); 318 continue; 319 } 320 321 /* do I own it? I should */ 322 PCTFS_rvec_zero(v,a_m); 323 if (col==fo[start]) { 324 start++; 325 idex=PCTFS_ivec_linear_search(col, a_local2global, a_n); 326 if (idex!=-1) { 327 v[idex] = 1.0; j++; 328 } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"NOT FOUND!\n"); 329 } else { 330 idex=PCTFS_ivec_linear_search(col, a_local2global, a_m); 331 if (idex!=-1) v[idex] = 1.0; 332 } 333 334 /* perform u = A.v_l */ 335 PCTFS_rvec_zero(u,n); 336 do_matvec(xxt_handle->mvi,v,u); 337 338 /* uu = X^T.u_l (local portion) */ 339 /* technically only need to zero out first i entries */ 340 /* later turn this into an XXT_solve call ? */ 341 PCTFS_rvec_zero(uu,m); 342 x_ptr=x; 343 iptr = col_indices; 344 for (k=0; k<i; k++) { 345 off = *iptr++; 346 len = *iptr++; 347 ierr = PetscBLASIntCast(len,&dlen);CHKERRQ(ierr); 348 PetscStackCallBLAS("BLASdot",uu[k] = BLASdot_(&dlen,u+off,&i1,x_ptr,&i1)); 349 x_ptr+=len; 350 } 351 352 /* uu = X^T.u_l (comm portion) */ 353 ierr = PCTFS_ssgl_radd (uu, w, dim, stages);CHKERRQ(ierr); 354 355 /* z = X.uu */ 356 PCTFS_rvec_zero(z,n); 357 x_ptr=x; 358 iptr = col_indices; 359 for (k=0; k<i; k++) { 360 off = *iptr++; 361 len = *iptr++; 362 ierr = PetscBLASIntCast(len,&dlen);CHKERRQ(ierr); 363 PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&dlen,&uu[k],x_ptr,&i1,z+off,&i1)); 364 x_ptr+=len; 365 } 366 367 /* compute v_l = v_l - z */ 368 PCTFS_rvec_zero(v+a_n,a_m-a_n); 369 ierr = PetscBLASIntCast(n,&dlen);CHKERRQ(ierr); 370 PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&dlen,&dm1,z,&i1,v,&i1)); 371 372 /* compute u_l = A.v_l */ 373 if (a_n!=a_m) PCTFS_gs_gop_hc(PCTFS_gs_handle,v,"+\0",dim); 374 PCTFS_rvec_zero(u,n); 375 do_matvec(xxt_handle->mvi,v,u); 376 377 /* compute sqrt(alpha) = sqrt(v_l^T.u_l) - local portion */ 378 ierr = PetscBLASIntCast(n,&dlen);CHKERRQ(ierr); 379 PetscStackCallBLAS("BLASdot",alpha = BLASdot_(&dlen,u,&i1,v,&i1)); 380 /* compute sqrt(alpha) = sqrt(v_l^T.u_l) - comm portion */ 381 PCTFS_grop_hc(&alpha, &alpha_w, 1, op, dim); 382 383 alpha = (PetscScalar) PetscSqrtReal((PetscReal)alpha); 384 385 /* check for small alpha */ 386 /* LATER use this to detect and determine null space */ 387 if (PetscAbsScalar(alpha)<1.0e-14) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_PLIB,"bad alpha! %g\n",alpha); 388 389 /* compute v_l = v_l/sqrt(alpha) */ 390 PCTFS_rvec_scale(v,1.0/alpha,n); 391 392 /* add newly generated column, v_l, to X */ 393 flag = 1; 394 off=len=0; 395 for (k=0; k<n; k++) { 396 if (v[k]!=0.0) { 397 len=k; 398 if (flag) { off=k; flag=0; } 399 } 400 } 401 402 len -= (off-1); 403 404 if (len>0) { 405 if ((xxt_nnz+len)>xxt_max_nnz) { 406 ierr = PetscInfo(0,"increasing space for X by 2x!\n");CHKERRQ(ierr); 407 xxt_max_nnz *= 2; 408 x_ptr = (PetscScalar*) malloc(xxt_max_nnz*sizeof(PetscScalar)); 409 PCTFS_rvec_copy(x_ptr,x,xxt_nnz); 410 free(x); 411 x = x_ptr; 412 x_ptr+=xxt_nnz; 413 } 414 xxt_nnz += len; 415 PCTFS_rvec_copy(x_ptr,v+off,len); 416 417 /* keep track of number of zeros */ 418 if (dim) { 419 for (k=0; k<len; k++) { 420 if (x_ptr[k]==0.0) xxt_zero_nnz++; 421 } 422 } else { 423 for (k=0; k<len; k++) { 424 if (x_ptr[k]==0.0) xxt_zero_nnz_0++; 425 } 426 } 427 col_indices[2*i] = off; 428 col_sz[i] = col_indices[2*i+1] = len; 429 col_vals[i] = x_ptr; 430 } 431 else { 432 col_indices[2*i] = 0; 433 col_sz[i] = col_indices[2*i+1] = 0; 434 col_vals[i] = x_ptr; 435 } 436 } 437 438 /* close off stages for execution phase */ 439 while (dim!=level) { 440 stages[dim++] = i; 441 ierr = PetscInfo2(0,"disconnected!!! dim(%D)!=level(%D)\n",dim,level);CHKERRQ(ierr); 442 } 443 stages[dim]=i; 444 445 xxt_handle->info->n = xxt_handle->mvi->n; 446 xxt_handle->info->m = m; 447 xxt_handle->info->nnz = xxt_nnz; 448 xxt_handle->info->max_nnz = xxt_max_nnz; 449 xxt_handle->info->msg_buf_sz = stages[level]-stages[0]; 450 xxt_handle->info->solve_uu = (PetscScalar*) malloc(m*sizeof(PetscScalar)); 451 xxt_handle->info->solve_w = (PetscScalar*) malloc(m*sizeof(PetscScalar)); 452 xxt_handle->info->x = x; 453 xxt_handle->info->col_vals = col_vals; 454 xxt_handle->info->col_sz = col_sz; 455 xxt_handle->info->col_indices = col_indices; 456 xxt_handle->info->stages = stages; 457 xxt_handle->info->nsolves = 0; 458 xxt_handle->info->tot_solve_time = 0.0; 459 460 free(segs); 461 free(u); 462 free(v); 463 free(uu); 464 free(z); 465 free(w); 466 467 return(0); 468 } 469 470 /**************************************xxt.c***********************************/ 471 static PetscErrorCode do_xxt_solve(xxt_ADT xxt_handle, PetscScalar *uc) 472 { 473 PetscInt off, len, *iptr; 474 PetscInt level = xxt_handle->level; 475 PetscInt n = xxt_handle->info->n; 476 PetscInt m = xxt_handle->info->m; 477 PetscInt *stages = xxt_handle->info->stages; 478 PetscInt *col_indices = xxt_handle->info->col_indices; 479 PetscScalar *x_ptr, *uu_ptr; 480 PetscScalar *solve_uu = xxt_handle->info->solve_uu; 481 PetscScalar *solve_w = xxt_handle->info->solve_w; 482 PetscScalar *x = xxt_handle->info->x; 483 PetscBLASInt i1 = 1,dlen; 484 PetscErrorCode ierr; 485 486 PetscFunctionBegin; 487 uu_ptr=solve_uu; 488 PCTFS_rvec_zero(uu_ptr,m); 489 490 /* x = X.Y^T.b */ 491 /* uu = Y^T.b */ 492 for (x_ptr=x,iptr=col_indices; *iptr!=-1; x_ptr+=len) { 493 off =*iptr++; 494 len =*iptr++; 495 ierr = PetscBLASIntCast(len,&dlen);CHKERRQ(ierr); 496 PetscStackCallBLAS("BLASdot",*uu_ptr++ = BLASdot_(&dlen,uc+off,&i1,x_ptr,&i1)); 497 } 498 499 /* comunication of beta */ 500 uu_ptr=solve_uu; 501 if (level) {ierr = PCTFS_ssgl_radd(uu_ptr, solve_w, level, stages);CHKERRQ(ierr);} 502 503 PCTFS_rvec_zero(uc,n); 504 505 /* x = X.uu */ 506 for (x_ptr=x,iptr=col_indices; *iptr!=-1; x_ptr+=len) { 507 off =*iptr++; 508 len =*iptr++; 509 ierr = PetscBLASIntCast(len,&dlen);CHKERRQ(ierr); 510 PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&dlen,uu_ptr++,x_ptr,&i1,uc+off,&i1)); 511 } 512 PetscFunctionReturn(0); 513 } 514 515 /**************************************xxt.c***********************************/ 516 static PetscErrorCode check_handle(xxt_ADT xxt_handle) 517 { 518 PetscInt vals[2], work[2], op[] = {NON_UNIFORM,GL_MIN,GL_MAX}; 519 520 PetscFunctionBegin; 521 if (!xxt_handle) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_PLIB,"check_handle() :: bad handle :: NULL %D\n",xxt_handle); 522 523 vals[0]=vals[1]=xxt_handle->id; 524 PCTFS_giop(vals,work,sizeof(op)/sizeof(op[0])-1,op); 525 if ((vals[0]!=vals[1])||(xxt_handle->id<=0)) SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_PLIB,"check_handle() :: bad handle :: id mismatch min/max %D/%D %D\n",vals[0],vals[1], xxt_handle->id); 526 PetscFunctionReturn(0); 527 } 528 529 /**************************************xxt.c***********************************/ 530 static PetscErrorCode det_separators(xxt_ADT xxt_handle) 531 { 532 PetscInt i, ct, id; 533 PetscInt mask, edge, *iptr; 534 PetscInt *dir, *used; 535 PetscInt sum[4], w[4]; 536 PetscScalar rsum[4], rw[4]; 537 PetscInt op[] = {GL_ADD,0}; 538 PetscScalar *lhs, *rhs; 539 PetscInt *nsep, *lnsep, *fo, nfo=0; 540 PCTFS_gs_ADT PCTFS_gs_handle = xxt_handle->mvi->PCTFS_gs_handle; 541 PetscInt *local2global = xxt_handle->mvi->local2global; 542 PetscInt n = xxt_handle->mvi->n; 543 PetscInt m = xxt_handle->mvi->m; 544 PetscInt level = xxt_handle->level; 545 PetscInt shared = 0; 546 547 PetscFunctionBegin; 548 dir = (PetscInt*)malloc(sizeof(PetscInt)*(level+1)); 549 nsep = (PetscInt*)malloc(sizeof(PetscInt)*(level+1)); 550 lnsep= (PetscInt*)malloc(sizeof(PetscInt)*(level+1)); 551 fo = (PetscInt*)malloc(sizeof(PetscInt)*(n+1)); 552 used = (PetscInt*)malloc(sizeof(PetscInt)*n); 553 554 PCTFS_ivec_zero(dir,level+1); 555 PCTFS_ivec_zero(nsep,level+1); 556 PCTFS_ivec_zero(lnsep,level+1); 557 PCTFS_ivec_set (fo,-1,n+1); 558 PCTFS_ivec_zero(used,n); 559 560 lhs = (PetscScalar*)malloc(sizeof(PetscScalar)*m); 561 rhs = (PetscScalar*)malloc(sizeof(PetscScalar)*m); 562 563 /* determine the # of unique dof */ 564 PCTFS_rvec_zero(lhs,m); 565 PCTFS_rvec_set(lhs,1.0,n); 566 PCTFS_gs_gop_hc(PCTFS_gs_handle,lhs,"+\0",level); 567 PCTFS_rvec_zero(rsum,2); 568 for (i=0; i<n; i++) { 569 if (lhs[i]!=0.0) { 570 rsum[0]+=1.0/lhs[i]; rsum[1]+=lhs[i]; 571 } 572 } 573 PCTFS_grop_hc(rsum,rw,2,op,level); 574 rsum[0]+=0.1; 575 rsum[1]+=0.1; 576 577 if (PetscAbsScalar(rsum[0]-rsum[1])>EPS) shared=1; 578 579 xxt_handle->info->n_global=xxt_handle->info->m_global=(PetscInt) rsum[0]; 580 xxt_handle->mvi->n_global =xxt_handle->mvi->m_global =(PetscInt) rsum[0]; 581 582 /* determine separator sets top down */ 583 if (shared) { 584 for (iptr=fo+n,id=PCTFS_my_id,mask=PCTFS_num_nodes>>1,edge=level;edge>0;edge--,mask>>=1) { 585 586 /* set rsh of hc, fire, and collect lhs responses */ 587 (id<mask) ? PCTFS_rvec_zero(lhs,m) : PCTFS_rvec_set(lhs,1.0,m); 588 PCTFS_gs_gop_hc(PCTFS_gs_handle,lhs,"+\0",edge); 589 590 /* set lsh of hc, fire, and collect rhs responses */ 591 (id<mask) ? PCTFS_rvec_set(rhs,1.0,m) : PCTFS_rvec_zero(rhs,m); 592 PCTFS_gs_gop_hc(PCTFS_gs_handle,rhs,"+\0",edge); 593 594 for (i=0;i<n;i++) { 595 if (id< mask) { 596 if (lhs[i]!=0.0) lhs[i]=1.0; 597 } 598 if (id>=mask) { 599 if (rhs[i]!=0.0) rhs[i]=1.0; 600 } 601 } 602 603 if (id< mask) PCTFS_gs_gop_hc(PCTFS_gs_handle,lhs,"+\0",edge-1); 604 else PCTFS_gs_gop_hc(PCTFS_gs_handle,rhs,"+\0",edge-1); 605 606 /* count number of dofs I own that have signal and not in sep set */ 607 PCTFS_rvec_zero(rsum,4); 608 for (PCTFS_ivec_zero(sum,4),ct=i=0;i<n;i++) { 609 if (!used[i]) { 610 /* number of unmarked dofs on node */ 611 ct++; 612 /* number of dofs to be marked on lhs hc */ 613 if (id< mask) { 614 if (lhs[i]!=0.0) { sum[0]++; rsum[0]+=1.0/lhs[i]; } 615 } 616 /* number of dofs to be marked on rhs hc */ 617 if (id>=mask) { 618 if (rhs[i]!=0.0) { sum[1]++; rsum[1]+=1.0/rhs[i]; } 619 } 620 } 621 } 622 623 /* go for load balance - choose half with most unmarked dofs, bias LHS */ 624 (id<mask) ? (sum[2]=ct) : (sum[3]=ct); 625 (id<mask) ? (rsum[2]=ct) : (rsum[3]=ct); 626 PCTFS_giop_hc(sum,w,4,op,edge); 627 PCTFS_grop_hc(rsum,rw,4,op,edge); 628 rsum[0]+=0.1; rsum[1]+=0.1; rsum[2]+=0.1; rsum[3]+=0.1; 629 630 if (id<mask) { 631 /* mark dofs I own that have signal and not in sep set */ 632 for (ct=i=0;i<n;i++) { 633 if ((!used[i])&&(lhs[i]!=0.0)) { 634 ct++; nfo++; 635 636 if (nfo>n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"nfo about to exceed n\n"); 637 638 *--iptr = local2global[i]; 639 used[i] = edge; 640 } 641 } 642 if (ct>1) PCTFS_ivec_sort(iptr,ct); 643 644 lnsep[edge]=ct; 645 nsep[edge]=(PetscInt) rsum[0]; 646 dir [edge]=LEFT; 647 } 648 649 if (id>=mask) { 650 /* mark dofs I own that have signal and not in sep set */ 651 for (ct=i=0;i<n;i++) { 652 if ((!used[i])&&(rhs[i]!=0.0)) { 653 ct++; nfo++; 654 655 if (nfo>n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"nfo about to exceed n\n"); 656 657 *--iptr = local2global[i]; 658 used[i] = edge; 659 } 660 } 661 if (ct>1) PCTFS_ivec_sort(iptr,ct); 662 663 lnsep[edge] = ct; 664 nsep[edge] = (PetscInt) rsum[1]; 665 dir [edge] = RIGHT; 666 } 667 668 /* LATER or we can recur on these to order seps at this level */ 669 /* do we need full set of separators for this? */ 670 671 /* fold rhs hc into lower */ 672 if (id>=mask) id-=mask; 673 } 674 } else { 675 for (iptr=fo+n,id=PCTFS_my_id,mask=PCTFS_num_nodes>>1,edge=level;edge>0;edge--,mask>>=1) { 676 /* set rsh of hc, fire, and collect lhs responses */ 677 (id<mask) ? PCTFS_rvec_zero(lhs,m) : PCTFS_rvec_set(lhs,1.0,m); 678 PCTFS_gs_gop_hc(PCTFS_gs_handle,lhs,"+\0",edge); 679 680 /* set lsh of hc, fire, and collect rhs responses */ 681 (id<mask) ? PCTFS_rvec_set(rhs,1.0,m) : PCTFS_rvec_zero(rhs,m); 682 PCTFS_gs_gop_hc(PCTFS_gs_handle,rhs,"+\0",edge); 683 684 /* count number of dofs I own that have signal and not in sep set */ 685 for (PCTFS_ivec_zero(sum,4),ct=i=0;i<n;i++) { 686 if (!used[i]) { 687 /* number of unmarked dofs on node */ 688 ct++; 689 /* number of dofs to be marked on lhs hc */ 690 if ((id< mask)&&(lhs[i]!=0.0)) sum[0]++; 691 /* number of dofs to be marked on rhs hc */ 692 if ((id>=mask)&&(rhs[i]!=0.0)) sum[1]++; 693 } 694 } 695 696 /* go for load balance - choose half with most unmarked dofs, bias LHS */ 697 (id<mask) ? (sum[2]=ct) : (sum[3]=ct); 698 PCTFS_giop_hc(sum,w,4,op,edge); 699 700 /* lhs hc wins */ 701 if (sum[2]>=sum[3]) { 702 if (id<mask) { 703 /* mark dofs I own that have signal and not in sep set */ 704 for (ct=i=0;i<n;i++) { 705 if ((!used[i])&&(lhs[i]!=0.0)) { 706 ct++; nfo++; 707 *--iptr = local2global[i]; 708 used[i]=edge; 709 } 710 } 711 if (ct>1) PCTFS_ivec_sort(iptr,ct); 712 lnsep[edge]=ct; 713 } 714 nsep[edge]=sum[0]; 715 dir [edge]=LEFT; 716 } else { /* rhs hc wins */ 717 if (id>=mask) { 718 /* mark dofs I own that have signal and not in sep set */ 719 for (ct=i=0;i<n;i++) { 720 if ((!used[i])&&(rhs[i]!=0.0)) { 721 ct++; nfo++; 722 *--iptr = local2global[i]; 723 used[i]=edge; 724 } 725 } 726 if (ct>1) PCTFS_ivec_sort(iptr,ct); 727 lnsep[edge]=ct; 728 } 729 nsep[edge]=sum[1]; 730 dir [edge]=RIGHT; 731 } 732 /* LATER or we can recur on these to order seps at this level */ 733 /* do we need full set of separators for this? */ 734 735 /* fold rhs hc into lower */ 736 if (id>=mask) id-=mask; 737 } 738 } 739 740 /* level 0 is on processor case - so mark the remainder */ 741 for (ct=i=0; i<n; i++) { 742 if (!used[i]) { 743 ct++; nfo++; 744 *--iptr = local2global[i]; 745 used[i] = edge; 746 } 747 } 748 if (ct>1) PCTFS_ivec_sort(iptr,ct); 749 lnsep[edge]=ct; 750 nsep [edge]=ct; 751 dir [edge]=LEFT; 752 753 xxt_handle->info->nsep = nsep; 754 xxt_handle->info->lnsep = lnsep; 755 xxt_handle->info->fo = fo; 756 xxt_handle->info->nfo = nfo; 757 758 free(dir); 759 free(lhs); 760 free(rhs); 761 free(used); 762 PetscFunctionReturn(0); 763 } 764 765 /**************************************xxt.c***********************************/ 766 static mv_info *set_mvi(PetscInt *local2global,PetscInt n,PetscInt m,PetscErrorCode (*matvec)(mv_info*,PetscScalar*,PetscScalar*),void *grid_data) 767 { 768 mv_info *mvi; 769 770 mvi = (mv_info*)malloc(sizeof(mv_info)); 771 mvi->n = n; 772 mvi->m = m; 773 mvi->n_global = -1; 774 mvi->m_global = -1; 775 mvi->local2global = (PetscInt*)malloc((m+1)*sizeof(PetscInt)); 776 PCTFS_ivec_copy(mvi->local2global,local2global,m); 777 mvi->local2global[m] = INT_MAX; 778 mvi->matvec = matvec; 779 mvi->grid_data = grid_data; 780 781 /* set xxt communication handle to perform restricted matvec */ 782 mvi->PCTFS_gs_handle = PCTFS_gs_init(local2global, m, PCTFS_num_nodes); 783 784 return(mvi); 785 } 786 787 /**************************************xxt.c***********************************/ 788 static PetscErrorCode do_matvec(mv_info *A, PetscScalar *v, PetscScalar *u) 789 { 790 PetscFunctionBegin; 791 A->matvec((mv_info*)A->grid_data,v,u); 792 PetscFunctionReturn(0); 793 } 794