1dba47a55SKris Buschelman #define PETSCKSP_DLL 2827bd09bSSatish Balay 3827bd09bSSatish Balay /*************************************xxt.c************************************ 4827bd09bSSatish Balay Module Name: xxt 5827bd09bSSatish Balay Module Info: 6827bd09bSSatish Balay 7827bd09bSSatish Balay author: Henry M. Tufo III 8827bd09bSSatish Balay e-mail: hmt@asci.uchicago.edu 9827bd09bSSatish Balay contact: 10827bd09bSSatish Balay +--------------------------------+--------------------------------+ 11827bd09bSSatish Balay |MCS Division - Building 221 |Department of Computer Science | 12827bd09bSSatish Balay |Argonne National Laboratory |Ryerson 152 | 13827bd09bSSatish Balay |9700 S. Cass Avenue |The University of Chicago | 14827bd09bSSatish Balay |Argonne, IL 60439 |Chicago, IL 60637 | 15827bd09bSSatish Balay |(630) 252-5354/5986 ph/fx |(773) 702-6019/8487 ph/fx | 16827bd09bSSatish Balay +--------------------------------+--------------------------------+ 17827bd09bSSatish Balay 18827bd09bSSatish Balay Last Modification: 3.20.01 19827bd09bSSatish Balay **************************************xxt.c***********************************/ 207c4f633dSBarry Smith #include "../src/ksp/pc/impls/tfs/tfs.h" 21827bd09bSSatish Balay 22827bd09bSSatish Balay #define LEFT -1 23827bd09bSSatish Balay #define RIGHT 1 24827bd09bSSatish Balay #define BOTH 0 25827bd09bSSatish Balay 26827bd09bSSatish Balay typedef struct xxt_solver_info { 2752f87cdaSBarry Smith PetscInt n, m, n_global, m_global; 2852f87cdaSBarry Smith PetscInt nnz, max_nnz, msg_buf_sz; 2952f87cdaSBarry Smith PetscInt *nsep, *lnsep, *fo, nfo, *stages; 3052f87cdaSBarry Smith PetscInt *col_sz, *col_indices; 31a501084fSBarry Smith PetscScalar **col_vals, *x, *solve_uu, *solve_w; 3252f87cdaSBarry Smith PetscInt nsolves; 33a501084fSBarry Smith PetscScalar tot_solve_time; 34827bd09bSSatish Balay } xxt_info; 35827bd09bSSatish Balay 36827bd09bSSatish Balay typedef struct matvec_info { 3752f87cdaSBarry Smith PetscInt n, m, n_global, m_global; 3852f87cdaSBarry Smith PetscInt *local2global; 39827bd09bSSatish Balay gs_ADT gs_handle; 40a501084fSBarry Smith PetscErrorCode (*matvec)(struct matvec_info*,PetscScalar*,PetscScalar*); 41827bd09bSSatish Balay void *grid_data; 42827bd09bSSatish Balay } mv_info; 43827bd09bSSatish Balay 44827bd09bSSatish Balay struct xxt_CDT{ 4552f87cdaSBarry Smith PetscInt id; 4652f87cdaSBarry Smith PetscInt ns; 4752f87cdaSBarry Smith PetscInt level; 48827bd09bSSatish Balay xxt_info *info; 49827bd09bSSatish Balay mv_info *mvi; 50827bd09bSSatish Balay }; 51827bd09bSSatish Balay 5252f87cdaSBarry Smith static PetscInt n_xxt=0; 5352f87cdaSBarry Smith static PetscInt n_xxt_handles=0; 54827bd09bSSatish Balay 55827bd09bSSatish Balay /* prototypes */ 563fdc5746SBarry Smith static PetscErrorCode do_xxt_solve(xxt_ADT xxt_handle, PetscScalar *rhs); 573fdc5746SBarry Smith static PetscErrorCode check_handle(xxt_ADT xxt_handle); 583fdc5746SBarry Smith static PetscErrorCode det_separators(xxt_ADT xxt_handle); 593fdc5746SBarry Smith static PetscErrorCode do_matvec(mv_info *A, PetscScalar *v, PetscScalar *u); 6052f87cdaSBarry Smith static PetscInt xxt_generate(xxt_ADT xxt_handle); 6152f87cdaSBarry Smith static PetscInt do_xxt_factor(xxt_ADT xxt_handle); 6252f87cdaSBarry Smith static mv_info *set_mvi(PetscInt *local2global, PetscInt n, PetscInt m, void *matvec, void *grid_data); 63a501084fSBarry Smith 647b1ae94cSBarry Smith /**************************************xxt.c***********************************/ 657b1ae94cSBarry Smith xxt_ADT XXT_new(void) 66827bd09bSSatish Balay { 67827bd09bSSatish Balay xxt_ADT xxt_handle; 68827bd09bSSatish Balay 69827bd09bSSatish Balay /* rolling count on n_xxt ... pot. problem here */ 70827bd09bSSatish Balay n_xxt_handles++; 71a501084fSBarry Smith xxt_handle = (xxt_ADT)malloc(sizeof(struct xxt_CDT)); 72827bd09bSSatish Balay xxt_handle->id = ++n_xxt; 73827bd09bSSatish Balay xxt_handle->info = NULL; xxt_handle->mvi = NULL; 74827bd09bSSatish Balay 75827bd09bSSatish Balay return(xxt_handle); 76827bd09bSSatish Balay } 77827bd09bSSatish Balay 787b1ae94cSBarry Smith /**************************************xxt.c***********************************/ 7952f87cdaSBarry Smith PetscInt XXT_factor(xxt_ADT xxt_handle, /* prev. allocated xxt handle */ 8052f87cdaSBarry Smith PetscInt *local2global, /* global column mapping */ 8152f87cdaSBarry Smith PetscInt n, /* local num rows */ 8252f87cdaSBarry Smith PetscInt m, /* local num cols */ 83827bd09bSSatish Balay void *matvec, /* b_loc=A_local.x_loc */ 84827bd09bSSatish Balay void *grid_data /* grid data for matvec */ 85827bd09bSSatish Balay ) 86827bd09bSSatish Balay { 87a40341efSBarry Smith comm_init(); 88827bd09bSSatish Balay check_handle(xxt_handle); 89827bd09bSSatish Balay 90827bd09bSSatish Balay /* only 2^k for now and all nodes participating */ 91*c1235816SBarry Smith if ((1<<(xxt_handle->level=i_log2_num_nodes))!=num_nodes) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"only 2^k for now and MPI_COMM_WORLD!!! %D != %D\n",1<<i_log2_num_nodes,num_nodes); 92827bd09bSSatish Balay 93827bd09bSSatish Balay /* space for X info */ 94a501084fSBarry Smith xxt_handle->info = (xxt_info*)malloc(sizeof(xxt_info)); 95827bd09bSSatish Balay 96827bd09bSSatish Balay /* set up matvec handles */ 97827bd09bSSatish Balay xxt_handle->mvi = set_mvi(local2global, n, m, matvec, grid_data); 98827bd09bSSatish Balay 99827bd09bSSatish Balay /* matrix is assumed to be of full rank */ 100827bd09bSSatish Balay /* LATER we can reset to indicate rank def. */ 101827bd09bSSatish Balay xxt_handle->ns=0; 102827bd09bSSatish Balay 103827bd09bSSatish Balay /* determine separators and generate firing order - NB xxt info set here */ 104827bd09bSSatish Balay det_separators(xxt_handle); 105827bd09bSSatish Balay 106827bd09bSSatish Balay return(do_xxt_factor(xxt_handle)); 107827bd09bSSatish Balay } 108827bd09bSSatish Balay 1097b1ae94cSBarry Smith /**************************************xxt.c***********************************/ 1108cda6cd7SBarry Smith PetscInt XXT_solve(xxt_ADT xxt_handle, PetscScalar *x, PetscScalar *b) 111827bd09bSSatish Balay { 112827bd09bSSatish Balay 113a40341efSBarry Smith comm_init(); 114827bd09bSSatish Balay check_handle(xxt_handle); 115827bd09bSSatish Balay 116827bd09bSSatish Balay /* need to copy b into x? */ 117827bd09bSSatish Balay if (b) 118827bd09bSSatish Balay {rvec_copy(x,b,xxt_handle->mvi->n);} 119827bd09bSSatish Balay do_xxt_solve(xxt_handle,x); 120827bd09bSSatish Balay 121827bd09bSSatish Balay return(0); 122827bd09bSSatish Balay } 123827bd09bSSatish Balay 1247b1ae94cSBarry Smith /**************************************xxt.c***********************************/ 12552f87cdaSBarry Smith PetscInt XXT_free(xxt_ADT xxt_handle) 126827bd09bSSatish Balay { 127827bd09bSSatish Balay 128a40341efSBarry Smith comm_init(); 129827bd09bSSatish Balay check_handle(xxt_handle); 130827bd09bSSatish Balay n_xxt_handles--; 131827bd09bSSatish Balay 132a501084fSBarry Smith free(xxt_handle->info->nsep); 133a501084fSBarry Smith free(xxt_handle->info->lnsep); 134a501084fSBarry Smith free(xxt_handle->info->fo); 135a501084fSBarry Smith free(xxt_handle->info->stages); 136a501084fSBarry Smith free(xxt_handle->info->solve_uu); 137a501084fSBarry Smith free(xxt_handle->info->solve_w); 138a501084fSBarry Smith free(xxt_handle->info->x); 139a501084fSBarry Smith free(xxt_handle->info->col_vals); 140a501084fSBarry Smith free(xxt_handle->info->col_sz); 141a501084fSBarry Smith free(xxt_handle->info->col_indices); 142a501084fSBarry Smith free(xxt_handle->info); 143a501084fSBarry Smith free(xxt_handle->mvi->local2global); 144827bd09bSSatish Balay gs_free(xxt_handle->mvi->gs_handle); 145a501084fSBarry Smith free(xxt_handle->mvi); 146a501084fSBarry Smith free(xxt_handle); 147827bd09bSSatish Balay 148827bd09bSSatish Balay /* if the check fails we nuke */ 149a501084fSBarry Smith /* if NULL pointer passed to free we nuke */ 150827bd09bSSatish Balay /* if the calls to free fail that's not my problem */ 151827bd09bSSatish Balay return(0); 152827bd09bSSatish Balay } 153827bd09bSSatish Balay 1547b1ae94cSBarry Smith /**************************************xxt.c***********************************/ 15552f87cdaSBarry Smith PetscInt XXT_stats(xxt_ADT xxt_handle) 156827bd09bSSatish Balay { 15752f87cdaSBarry Smith PetscInt op[] = {NON_UNIFORM,GL_MIN,GL_MAX,GL_ADD,GL_MIN,GL_MAX,GL_ADD,GL_MIN,GL_MAX,GL_ADD}; 15852f87cdaSBarry Smith PetscInt fop[] = {NON_UNIFORM,GL_MIN,GL_MAX,GL_ADD}; 15952f87cdaSBarry Smith PetscInt vals[9], work[9]; 160a501084fSBarry Smith PetscScalar fvals[3], fwork[3]; 16171a0148aSBarry Smith PetscErrorCode ierr; 162827bd09bSSatish Balay 163a40341efSBarry Smith comm_init(); 164827bd09bSSatish Balay check_handle(xxt_handle); 165827bd09bSSatish Balay 166827bd09bSSatish Balay /* if factorization not done there are no stats */ 167827bd09bSSatish Balay if (!xxt_handle->info||!xxt_handle->mvi) 168827bd09bSSatish Balay { 169827bd09bSSatish Balay if (!my_id) 17071a0148aSBarry Smith {ierr = PetscPrintf(PETSC_COMM_WORLD,"XXT_stats() :: no stats available!\n");} 171827bd09bSSatish Balay return 1; 172827bd09bSSatish Balay } 173827bd09bSSatish Balay 174827bd09bSSatish Balay vals[0]=vals[1]=vals[2]=xxt_handle->info->nnz; 175827bd09bSSatish Balay vals[3]=vals[4]=vals[5]=xxt_handle->mvi->n; 176827bd09bSSatish Balay vals[6]=vals[7]=vals[8]=xxt_handle->info->msg_buf_sz; 177827bd09bSSatish Balay giop(vals,work,sizeof(op)/sizeof(op[0])-1,op); 178827bd09bSSatish Balay 179827bd09bSSatish Balay fvals[0]=fvals[1]=fvals[2] 180827bd09bSSatish Balay =xxt_handle->info->tot_solve_time/xxt_handle->info->nsolves++; 181827bd09bSSatish Balay grop(fvals,fwork,sizeof(fop)/sizeof(fop[0])-1,fop); 182827bd09bSSatish Balay 183827bd09bSSatish Balay if (!my_id) 184827bd09bSSatish Balay { 18571a0148aSBarry Smith ierr = PetscPrintf(PETSC_COMM_WORLD,"%D :: min xxt_nnz=%D\n",my_id,vals[0]); 18671a0148aSBarry Smith ierr = PetscPrintf(PETSC_COMM_WORLD,"%D :: max xxt_nnz=%D\n",my_id,vals[1]); 18771a0148aSBarry Smith ierr = PetscPrintf(PETSC_COMM_WORLD,"%D :: avg xxt_nnz=%g\n",my_id,1.0*vals[2]/num_nodes); 18871a0148aSBarry Smith ierr = PetscPrintf(PETSC_COMM_WORLD,"%D :: tot xxt_nnz=%D\n",my_id,vals[2]); 18971a0148aSBarry Smith ierr = PetscPrintf(PETSC_COMM_WORLD,"%D :: xxt C(2d) =%g\n",my_id,vals[2]/(pow(1.0*vals[5],1.5))); 19071a0148aSBarry Smith ierr = PetscPrintf(PETSC_COMM_WORLD,"%D :: xxt C(3d) =%g\n",my_id,vals[2]/(pow(1.0*vals[5],1.6667))); 19171a0148aSBarry Smith ierr = PetscPrintf(PETSC_COMM_WORLD,"%D :: min xxt_n =%D\n",my_id,vals[3]); 19271a0148aSBarry Smith ierr = PetscPrintf(PETSC_COMM_WORLD,"%D :: max xxt_n =%D\n",my_id,vals[4]); 19371a0148aSBarry Smith ierr = PetscPrintf(PETSC_COMM_WORLD,"%D :: avg xxt_n =%g\n",my_id,1.0*vals[5]/num_nodes); 19471a0148aSBarry Smith ierr = PetscPrintf(PETSC_COMM_WORLD,"%D :: tot xxt_n =%D\n",my_id,vals[5]); 19571a0148aSBarry Smith ierr = PetscPrintf(PETSC_COMM_WORLD,"%D :: min xxt_buf=%D\n",my_id,vals[6]); 19671a0148aSBarry Smith ierr = PetscPrintf(PETSC_COMM_WORLD,"%D :: max xxt_buf=%D\n",my_id,vals[7]); 19771a0148aSBarry Smith ierr = PetscPrintf(PETSC_COMM_WORLD,"%D :: avg xxt_buf=%g\n",my_id,1.0*vals[8]/num_nodes); 19871a0148aSBarry Smith ierr = PetscPrintf(PETSC_COMM_WORLD,"%D :: min xxt_slv=%g\n",my_id,fvals[0]); 19971a0148aSBarry Smith ierr = PetscPrintf(PETSC_COMM_WORLD,"%D :: max xxt_slv=%g\n",my_id,fvals[1]); 20071a0148aSBarry Smith ierr = PetscPrintf(PETSC_COMM_WORLD,"%D :: avg xxt_slv=%g\n",my_id,fvals[2]/num_nodes); 201827bd09bSSatish Balay } 202827bd09bSSatish Balay 203827bd09bSSatish Balay return(0); 204827bd09bSSatish Balay } 205827bd09bSSatish Balay 206827bd09bSSatish Balay /*************************************xxt.c************************************ 207827bd09bSSatish Balay 208827bd09bSSatish Balay Description: get A_local, local portion of global coarse matrix which 209827bd09bSSatish Balay is a row dist. nxm matrix w/ n<m. 210827bd09bSSatish Balay o my_ml holds address of ML struct associated w/A_local and coarse grid 211827bd09bSSatish Balay o local2global holds global number of column i (i=0,...,m-1) 212827bd09bSSatish Balay o local2global holds global number of row i (i=0,...,n-1) 213827bd09bSSatish Balay o mylocmatvec performs A_local . vec_local (note that gs is performed using 214827bd09bSSatish Balay gs_init/gop). 215827bd09bSSatish Balay 216827bd09bSSatish Balay mylocmatvec = my_ml->Amat[grid_tag].matvec->external; 217827bd09bSSatish Balay mylocmatvec (void :: void *data, double *in, double *out) 218827bd09bSSatish Balay **************************************xxt.c***********************************/ 21952f87cdaSBarry Smith static PetscInt do_xxt_factor(xxt_ADT xxt_handle) 220827bd09bSSatish Balay { 2217b1ae94cSBarry Smith return xxt_generate(xxt_handle); 222827bd09bSSatish Balay } 223827bd09bSSatish Balay 2247b1ae94cSBarry Smith /**************************************xxt.c***********************************/ 22552f87cdaSBarry Smith static PetscInt xxt_generate(xxt_ADT xxt_handle) 226827bd09bSSatish Balay { 22752f87cdaSBarry Smith PetscInt i,j,k,idex; 22852f87cdaSBarry Smith PetscInt dim, col; 229a501084fSBarry Smith PetscScalar *u, *uu, *v, *z, *w, alpha, alpha_w; 23052f87cdaSBarry Smith PetscInt *segs; 23152f87cdaSBarry Smith PetscInt op[] = {GL_ADD,0}; 23252f87cdaSBarry Smith PetscInt off, len; 233a501084fSBarry Smith PetscScalar *x_ptr; 23452f87cdaSBarry Smith PetscInt *iptr, flag; 23552f87cdaSBarry Smith PetscInt start=0, end, work; 23652f87cdaSBarry Smith PetscInt op2[] = {GL_MIN,0}; 237827bd09bSSatish Balay gs_ADT gs_handle; 23852f87cdaSBarry Smith PetscInt *nsep, *lnsep, *fo; 23952f87cdaSBarry Smith PetscInt a_n=xxt_handle->mvi->n; 24052f87cdaSBarry Smith PetscInt a_m=xxt_handle->mvi->m; 24152f87cdaSBarry Smith PetscInt *a_local2global=xxt_handle->mvi->local2global; 24252f87cdaSBarry Smith PetscInt level; 24352f87cdaSBarry Smith PetscInt xxt_nnz=0, xxt_max_nnz=0; 24452f87cdaSBarry Smith PetscInt n, m; 24552f87cdaSBarry Smith PetscInt *col_sz, *col_indices, *stages; 246a501084fSBarry Smith PetscScalar **col_vals, *x; 24752f87cdaSBarry Smith PetscInt n_global; 24852f87cdaSBarry Smith PetscInt xxt_zero_nnz=0; 24952f87cdaSBarry Smith PetscInt xxt_zero_nnz_0=0; 25071a0148aSBarry Smith PetscBLASInt i1 = 1,dlen; 251a501084fSBarry Smith PetscScalar dm1 = -1.0; 252d1528f56SBarry Smith PetscErrorCode ierr; 253827bd09bSSatish Balay 254827bd09bSSatish Balay n=xxt_handle->mvi->n; 255827bd09bSSatish Balay nsep=xxt_handle->info->nsep; 256827bd09bSSatish Balay lnsep=xxt_handle->info->lnsep; 257827bd09bSSatish Balay fo=xxt_handle->info->fo; 258827bd09bSSatish Balay end=lnsep[0]; 259827bd09bSSatish Balay level=xxt_handle->level; 260827bd09bSSatish Balay gs_handle=xxt_handle->mvi->gs_handle; 261827bd09bSSatish Balay 262827bd09bSSatish Balay /* is there a null space? */ 263827bd09bSSatish Balay /* LATER add in ability to detect null space by checking alpha */ 264827bd09bSSatish Balay for (i=0, j=0; i<=level; i++) 265827bd09bSSatish Balay {j+=nsep[i];} 266827bd09bSSatish Balay 267827bd09bSSatish Balay m = j-xxt_handle->ns; 268827bd09bSSatish Balay if (m!=j) 26971a0148aSBarry Smith {ierr = PetscPrintf(PETSC_COMM_WORLD,"xxt_generate() :: null space exists %D %D %D\n",m,j,xxt_handle->ns);} 270827bd09bSSatish Balay 271827bd09bSSatish Balay /* get and initialize storage for x local */ 272827bd09bSSatish Balay /* note that x local is nxm and stored by columns */ 27352f87cdaSBarry Smith col_sz = (PetscInt*) malloc(m*sizeof(PetscInt)); 27452f87cdaSBarry Smith col_indices = (PetscInt*) malloc((2*m+1)*sizeof(PetscInt)); 275a501084fSBarry Smith col_vals = (PetscScalar **) malloc(m*sizeof(PetscScalar *)); 276827bd09bSSatish Balay for (i=j=0; i<m; i++, j+=2) 277827bd09bSSatish Balay { 278827bd09bSSatish Balay col_indices[j]=col_indices[j+1]=col_sz[i]=-1; 279827bd09bSSatish Balay col_vals[i] = NULL; 280827bd09bSSatish Balay } 281827bd09bSSatish Balay col_indices[j]=-1; 282827bd09bSSatish Balay 283827bd09bSSatish Balay /* size of separators for each sub-hc working from bottom of tree to top */ 284827bd09bSSatish Balay /* this looks like nsep[]=segments */ 28552f87cdaSBarry Smith stages = (PetscInt*) malloc((level+1)*sizeof(PetscInt)); 28652f87cdaSBarry Smith segs = (PetscInt*) malloc((level+1)*sizeof(PetscInt)); 287827bd09bSSatish Balay ivec_zero(stages,level+1); 288827bd09bSSatish Balay ivec_copy(segs,nsep,level+1); 289827bd09bSSatish Balay for (i=0; i<level; i++) 290827bd09bSSatish Balay {segs[i+1] += segs[i];} 291827bd09bSSatish Balay stages[0] = segs[0]; 292827bd09bSSatish Balay 293827bd09bSSatish Balay /* temporary vectors */ 294a501084fSBarry Smith u = (PetscScalar *) malloc(n*sizeof(PetscScalar)); 295a501084fSBarry Smith z = (PetscScalar *) malloc(n*sizeof(PetscScalar)); 296a501084fSBarry Smith v = (PetscScalar *) malloc(a_m*sizeof(PetscScalar)); 297a501084fSBarry Smith uu = (PetscScalar *) malloc(m*sizeof(PetscScalar)); 298a501084fSBarry Smith w = (PetscScalar *) malloc(m*sizeof(PetscScalar)); 299827bd09bSSatish Balay 300827bd09bSSatish Balay /* extra nnz due to replication of vertices across separators */ 301827bd09bSSatish Balay for (i=1, j=0; i<=level; i++) 302827bd09bSSatish Balay {j+=nsep[i];} 303827bd09bSSatish Balay 304827bd09bSSatish Balay /* storage for sparse x values */ 305827bd09bSSatish Balay n_global = xxt_handle->info->n_global; 30652f87cdaSBarry Smith xxt_max_nnz = (PetscInt)(2.5*pow(1.0*n_global,1.6667) + j*n/2)/num_nodes; 307a501084fSBarry Smith x = (PetscScalar *) malloc(xxt_max_nnz*sizeof(PetscScalar)); 308827bd09bSSatish Balay xxt_nnz = 0; 309827bd09bSSatish Balay 310827bd09bSSatish Balay /* LATER - can embed next sep to fire in gs */ 311827bd09bSSatish Balay /* time to make the donuts - generate X factor */ 312827bd09bSSatish Balay for (dim=i=j=0;i<m;i++) 313827bd09bSSatish Balay { 314827bd09bSSatish Balay /* time to move to the next level? */ 315d4af0d30SBarry Smith while (i==segs[dim]) { 316e32f2f54SBarry Smith if (dim==level) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"dim about to exceed level\n"); 317827bd09bSSatish Balay stages[dim++]=i; 318827bd09bSSatish Balay end+=lnsep[dim]; 319827bd09bSSatish Balay } 320827bd09bSSatish Balay stages[dim]=i; 321827bd09bSSatish Balay 322827bd09bSSatish Balay /* which column are we firing? */ 323827bd09bSSatish Balay /* i.e. set v_l */ 324827bd09bSSatish Balay /* use new seps and do global min across hc to determine which one to fire */ 325827bd09bSSatish Balay (start<end) ? (col=fo[start]) : (col=INT_MAX); 326827bd09bSSatish Balay giop_hc(&col,&work,1,op2,dim); 327827bd09bSSatish Balay 328827bd09bSSatish Balay /* shouldn't need this */ 329827bd09bSSatish Balay if (col==INT_MAX) 330827bd09bSSatish Balay { 331f1ed62a8SBarry Smith ierr = PetscInfo(0,"hey ... col==INT_MAX??\n");CHKERRQ(ierr); 332827bd09bSSatish Balay continue; 333827bd09bSSatish Balay } 334827bd09bSSatish Balay 335827bd09bSSatish Balay /* do I own it? I should */ 336827bd09bSSatish Balay rvec_zero(v ,a_m); 337827bd09bSSatish Balay if (col==fo[start]) 338827bd09bSSatish Balay { 339827bd09bSSatish Balay start++; 340827bd09bSSatish Balay idex=ivec_linear_search(col, a_local2global, a_n); 341e7e72b3dSBarry Smith if (idex!=-1){ 342e7e72b3dSBarry Smith v[idex] = 1.0; j++; 343e7e72b3dSBarry Smith } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"NOT FOUND!\n"); 344827bd09bSSatish Balay } 345827bd09bSSatish Balay else 346827bd09bSSatish Balay { 347827bd09bSSatish Balay idex=ivec_linear_search(col, a_local2global, a_m); 348827bd09bSSatish Balay if (idex!=-1) 349827bd09bSSatish Balay {v[idex] = 1.0;} 350827bd09bSSatish Balay } 351827bd09bSSatish Balay 352827bd09bSSatish Balay /* perform u = A.v_l */ 353827bd09bSSatish Balay rvec_zero(u,n); 354827bd09bSSatish Balay do_matvec(xxt_handle->mvi,v,u); 355827bd09bSSatish Balay 356827bd09bSSatish Balay /* uu = X^T.u_l (local portion) */ 357827bd09bSSatish Balay /* technically only need to zero out first i entries */ 358827bd09bSSatish Balay /* later turn this into an XXT_solve call ? */ 359827bd09bSSatish Balay rvec_zero(uu,m); 360827bd09bSSatish Balay x_ptr=x; 361827bd09bSSatish Balay iptr = col_indices; 362827bd09bSSatish Balay for (k=0; k<i; k++) 363827bd09bSSatish Balay { 364827bd09bSSatish Balay off = *iptr++; 365827bd09bSSatish Balay len = *iptr++; 3660805154bSBarry Smith dlen = PetscBLASIntCast(len); 36771a0148aSBarry Smith uu[k] = BLASdot_(&dlen,u+off,&i1,x_ptr,&i1); 368827bd09bSSatish Balay x_ptr+=len; 369827bd09bSSatish Balay } 370827bd09bSSatish Balay 371827bd09bSSatish Balay 372827bd09bSSatish Balay /* uu = X^T.u_l (comm portion) */ 373827bd09bSSatish Balay ssgl_radd (uu, w, dim, stages); 374827bd09bSSatish Balay 375827bd09bSSatish Balay /* z = X.uu */ 376827bd09bSSatish Balay rvec_zero(z,n); 377827bd09bSSatish Balay x_ptr=x; 378827bd09bSSatish Balay iptr = col_indices; 379827bd09bSSatish Balay for (k=0; k<i; k++) 380827bd09bSSatish Balay { 381827bd09bSSatish Balay off = *iptr++; 382827bd09bSSatish Balay len = *iptr++; 3830805154bSBarry Smith dlen = PetscBLASIntCast(len); 38471a0148aSBarry Smith BLASaxpy_(&dlen,&uu[k],x_ptr,&i1,z+off,&i1); 385827bd09bSSatish Balay x_ptr+=len; 386827bd09bSSatish Balay } 387827bd09bSSatish Balay 388827bd09bSSatish Balay /* compute v_l = v_l - z */ 389827bd09bSSatish Balay rvec_zero(v+a_n,a_m-a_n); 3900805154bSBarry Smith dlen = PetscBLASIntCast(n); 39171a0148aSBarry Smith BLASaxpy_(&dlen,&dm1,z,&i1,v,&i1); 392827bd09bSSatish Balay 393827bd09bSSatish Balay /* compute u_l = A.v_l */ 394827bd09bSSatish Balay if (a_n!=a_m) 395827bd09bSSatish Balay {gs_gop_hc(gs_handle,v,"+\0",dim);} 396827bd09bSSatish Balay rvec_zero(u,n); 397827bd09bSSatish Balay do_matvec(xxt_handle->mvi,v,u); 398827bd09bSSatish Balay 399827bd09bSSatish Balay /* compute sqrt(alpha) = sqrt(v_l^T.u_l) - local portion */ 4000805154bSBarry Smith dlen = PetscBLASIntCast(n); 40171a0148aSBarry Smith alpha = BLASdot_(&dlen,u,&i1,v,&i1); 402827bd09bSSatish Balay /* compute sqrt(alpha) = sqrt(v_l^T.u_l) - comm portion */ 403827bd09bSSatish Balay grop_hc(&alpha, &alpha_w, 1, op, dim); 404827bd09bSSatish Balay 405a501084fSBarry Smith alpha = (PetscScalar) sqrt((double)alpha); 406827bd09bSSatish Balay 407827bd09bSSatish Balay /* check for small alpha */ 408827bd09bSSatish Balay /* LATER use this to detect and determine null space */ 409*c1235816SBarry Smith if (fabs(alpha)<1.0e-14) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_PLIB,"bad alpha! %g\n",alpha); 410827bd09bSSatish Balay 411827bd09bSSatish Balay /* compute v_l = v_l/sqrt(alpha) */ 412827bd09bSSatish Balay rvec_scale(v,1.0/alpha,n); 413827bd09bSSatish Balay 414827bd09bSSatish Balay /* add newly generated column, v_l, to X */ 415827bd09bSSatish Balay flag = 1; 416827bd09bSSatish Balay off=len=0; 417827bd09bSSatish Balay for (k=0; k<n; k++) 418827bd09bSSatish Balay { 419827bd09bSSatish Balay if (v[k]!=0.0) 420827bd09bSSatish Balay { 421827bd09bSSatish Balay len=k; 422827bd09bSSatish Balay if (flag) 423827bd09bSSatish Balay {off=k; flag=0;} 424827bd09bSSatish Balay } 425827bd09bSSatish Balay } 426827bd09bSSatish Balay 427827bd09bSSatish Balay len -= (off-1); 428827bd09bSSatish Balay 429827bd09bSSatish Balay if (len>0) 430827bd09bSSatish Balay { 431827bd09bSSatish Balay if ((xxt_nnz+len)>xxt_max_nnz) 432827bd09bSSatish Balay { 433f1ed62a8SBarry Smith ierr = PetscInfo(0,"increasing space for X by 2x!\n");CHKERRQ(ierr); 434827bd09bSSatish Balay xxt_max_nnz *= 2; 435a501084fSBarry Smith x_ptr = (PetscScalar *) malloc(xxt_max_nnz*sizeof(PetscScalar)); 436827bd09bSSatish Balay rvec_copy(x_ptr,x,xxt_nnz); 437a501084fSBarry Smith free(x); 438827bd09bSSatish Balay x = x_ptr; 439827bd09bSSatish Balay x_ptr+=xxt_nnz; 440827bd09bSSatish Balay } 441827bd09bSSatish Balay xxt_nnz += len; 442827bd09bSSatish Balay rvec_copy(x_ptr,v+off,len); 443827bd09bSSatish Balay 444827bd09bSSatish Balay /* keep track of number of zeros */ 445827bd09bSSatish Balay if (dim) 446827bd09bSSatish Balay { 447827bd09bSSatish Balay for (k=0; k<len; k++) 448827bd09bSSatish Balay { 449827bd09bSSatish Balay if (x_ptr[k]==0.0) 450827bd09bSSatish Balay {xxt_zero_nnz++;} 451827bd09bSSatish Balay } 452827bd09bSSatish Balay } 453827bd09bSSatish Balay else 454827bd09bSSatish Balay { 455827bd09bSSatish Balay for (k=0; k<len; k++) 456827bd09bSSatish Balay { 457827bd09bSSatish Balay if (x_ptr[k]==0.0) 458827bd09bSSatish Balay {xxt_zero_nnz_0++;} 459827bd09bSSatish Balay } 460827bd09bSSatish Balay } 461827bd09bSSatish Balay col_indices[2*i] = off; 462827bd09bSSatish Balay col_sz[i] = col_indices[2*i+1] = len; 463827bd09bSSatish Balay col_vals[i] = x_ptr; 464827bd09bSSatish Balay } 465827bd09bSSatish Balay else 466827bd09bSSatish Balay { 467827bd09bSSatish Balay col_indices[2*i] = 0; 468827bd09bSSatish Balay col_sz[i] = col_indices[2*i+1] = 0; 469827bd09bSSatish Balay col_vals[i] = x_ptr; 470827bd09bSSatish Balay } 471827bd09bSSatish Balay } 472827bd09bSSatish Balay 473827bd09bSSatish Balay /* close off stages for execution phase */ 474827bd09bSSatish Balay while (dim!=level) 475827bd09bSSatish Balay { 476827bd09bSSatish Balay stages[dim++]=i; 47771a0148aSBarry Smith ierr = PetscInfo2(0,"disconnected!!! dim(%D)!=level(%D)\n",dim,level);CHKERRQ(ierr); 478827bd09bSSatish Balay } 479827bd09bSSatish Balay stages[dim]=i; 480827bd09bSSatish Balay 481827bd09bSSatish Balay xxt_handle->info->n=xxt_handle->mvi->n; 482827bd09bSSatish Balay xxt_handle->info->m=m; 483827bd09bSSatish Balay xxt_handle->info->nnz=xxt_nnz; 484827bd09bSSatish Balay xxt_handle->info->max_nnz=xxt_max_nnz; 485827bd09bSSatish Balay xxt_handle->info->msg_buf_sz=stages[level]-stages[0]; 486a501084fSBarry Smith xxt_handle->info->solve_uu = (PetscScalar *) malloc(m*sizeof(PetscScalar)); 487a501084fSBarry Smith xxt_handle->info->solve_w = (PetscScalar *) malloc(m*sizeof(PetscScalar)); 488827bd09bSSatish Balay xxt_handle->info->x=x; 489827bd09bSSatish Balay xxt_handle->info->col_vals=col_vals; 490827bd09bSSatish Balay xxt_handle->info->col_sz=col_sz; 491827bd09bSSatish Balay xxt_handle->info->col_indices=col_indices; 492827bd09bSSatish Balay xxt_handle->info->stages=stages; 493827bd09bSSatish Balay xxt_handle->info->nsolves=0; 494827bd09bSSatish Balay xxt_handle->info->tot_solve_time=0.0; 495827bd09bSSatish Balay 496a501084fSBarry Smith free(segs); 497a501084fSBarry Smith free(u); 498a501084fSBarry Smith free(v); 499a501084fSBarry Smith free(uu); 500a501084fSBarry Smith free(z); 501a501084fSBarry Smith free(w); 502827bd09bSSatish Balay 503827bd09bSSatish Balay return(0); 504827bd09bSSatish Balay } 505827bd09bSSatish Balay 5067b1ae94cSBarry Smith /**************************************xxt.c***********************************/ 5070924e98cSBarry Smith static PetscErrorCode do_xxt_solve(xxt_ADT xxt_handle, PetscScalar *uc) 508827bd09bSSatish Balay { 50952f87cdaSBarry Smith PetscInt off, len, *iptr; 51052f87cdaSBarry Smith PetscInt level =xxt_handle->level; 51152f87cdaSBarry Smith PetscInt n =xxt_handle->info->n; 51252f87cdaSBarry Smith PetscInt m =xxt_handle->info->m; 51352f87cdaSBarry Smith PetscInt *stages =xxt_handle->info->stages; 51452f87cdaSBarry Smith PetscInt *col_indices=xxt_handle->info->col_indices; 515a501084fSBarry Smith PetscScalar *x_ptr, *uu_ptr; 516a501084fSBarry Smith PetscScalar *solve_uu=xxt_handle->info->solve_uu; 517a501084fSBarry Smith PetscScalar *solve_w =xxt_handle->info->solve_w; 518a501084fSBarry Smith PetscScalar *x =xxt_handle->info->x; 51971a0148aSBarry Smith PetscBLASInt i1 = 1,dlen; 520827bd09bSSatish Balay 5210924e98cSBarry Smith PetscFunctionBegin; 522827bd09bSSatish Balay uu_ptr=solve_uu; 523827bd09bSSatish Balay rvec_zero(uu_ptr,m); 524827bd09bSSatish Balay 525827bd09bSSatish Balay /* x = X.Y^T.b */ 526827bd09bSSatish Balay /* uu = Y^T.b */ 527827bd09bSSatish Balay for (x_ptr=x,iptr=col_indices; *iptr!=-1; x_ptr+=len) 528827bd09bSSatish Balay { 529827bd09bSSatish Balay off=*iptr++; len=*iptr++; 5300805154bSBarry Smith dlen = PetscBLASIntCast(len); 53171a0148aSBarry Smith *uu_ptr++ = BLASdot_(&dlen,uc+off,&i1,x_ptr,&i1); 532827bd09bSSatish Balay } 533827bd09bSSatish Balay 534827bd09bSSatish Balay /* comunication of beta */ 535827bd09bSSatish Balay uu_ptr=solve_uu; 536827bd09bSSatish Balay if (level) {ssgl_radd(uu_ptr, solve_w, level, stages);} 537827bd09bSSatish Balay 538827bd09bSSatish Balay rvec_zero(uc,n); 539827bd09bSSatish Balay 540827bd09bSSatish Balay /* x = X.uu */ 541827bd09bSSatish Balay for (x_ptr=x,iptr=col_indices; *iptr!=-1; x_ptr+=len) 542827bd09bSSatish Balay { 543827bd09bSSatish Balay off=*iptr++; len=*iptr++; 5440805154bSBarry Smith dlen = PetscBLASIntCast(len); 54571a0148aSBarry Smith BLASaxpy_(&dlen,uu_ptr++,x_ptr,&i1,uc+off,&i1); 546827bd09bSSatish Balay } 5470924e98cSBarry Smith PetscFunctionReturn(0); 548827bd09bSSatish Balay } 549827bd09bSSatish Balay 5507b1ae94cSBarry Smith /**************************************xxt.c***********************************/ 5510924e98cSBarry Smith static PetscErrorCode check_handle(xxt_ADT xxt_handle) 552827bd09bSSatish Balay { 55352f87cdaSBarry Smith PetscInt vals[2], work[2], op[] = {NON_UNIFORM,GL_MIN,GL_MAX}; 554827bd09bSSatish Balay 5550924e98cSBarry Smith PetscFunctionBegin; 556e32f2f54SBarry Smith if (!xxt_handle) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_PLIB,"check_handle() :: bad handle :: NULL %D\n",xxt_handle); 557827bd09bSSatish Balay 558827bd09bSSatish Balay vals[0]=vals[1]=xxt_handle->id; 559827bd09bSSatish Balay giop(vals,work,sizeof(op)/sizeof(op[0])-1,op); 560e32f2f54SBarry Smith 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); 5610924e98cSBarry Smith PetscFunctionReturn(0); 562827bd09bSSatish Balay } 563827bd09bSSatish Balay 5647b1ae94cSBarry Smith /**************************************xxt.c***********************************/ 5650924e98cSBarry Smith static PetscErrorCode det_separators(xxt_ADT xxt_handle) 566827bd09bSSatish Balay { 56752f87cdaSBarry Smith PetscInt i, ct, id; 56852f87cdaSBarry Smith PetscInt mask, edge, *iptr; 56952f87cdaSBarry Smith PetscInt *dir, *used; 57052f87cdaSBarry Smith PetscInt sum[4], w[4]; 571a501084fSBarry Smith PetscScalar rsum[4], rw[4]; 57252f87cdaSBarry Smith PetscInt op[] = {GL_ADD,0}; 573a501084fSBarry Smith PetscScalar *lhs, *rhs; 57452f87cdaSBarry Smith PetscInt *nsep, *lnsep, *fo, nfo=0; 575827bd09bSSatish Balay gs_ADT gs_handle=xxt_handle->mvi->gs_handle; 57652f87cdaSBarry Smith PetscInt *local2global=xxt_handle->mvi->local2global; 57752f87cdaSBarry Smith PetscInt n=xxt_handle->mvi->n; 57852f87cdaSBarry Smith PetscInt m=xxt_handle->mvi->m; 57952f87cdaSBarry Smith PetscInt level=xxt_handle->level; 58052f87cdaSBarry Smith PetscInt shared=FALSE; 581827bd09bSSatish Balay 5820924e98cSBarry Smith PetscFunctionBegin; 58352f87cdaSBarry Smith dir = (PetscInt*)malloc(sizeof(PetscInt)*(level+1)); 58452f87cdaSBarry Smith nsep = (PetscInt*)malloc(sizeof(PetscInt)*(level+1)); 58552f87cdaSBarry Smith lnsep= (PetscInt*)malloc(sizeof(PetscInt)*(level+1)); 58652f87cdaSBarry Smith fo = (PetscInt*)malloc(sizeof(PetscInt)*(n+1)); 58752f87cdaSBarry Smith used = (PetscInt*)malloc(sizeof(PetscInt)*n); 588827bd09bSSatish Balay 589827bd09bSSatish Balay ivec_zero(dir ,level+1); 590827bd09bSSatish Balay ivec_zero(nsep ,level+1); 591827bd09bSSatish Balay ivec_zero(lnsep,level+1); 592827bd09bSSatish Balay ivec_set (fo ,-1,n+1); 593827bd09bSSatish Balay ivec_zero(used,n); 594827bd09bSSatish Balay 5958cda6cd7SBarry Smith lhs = (PetscScalar*)malloc(sizeof(PetscScalar)*m); 5968cda6cd7SBarry Smith rhs = (PetscScalar*)malloc(sizeof(PetscScalar)*m); 597827bd09bSSatish Balay 598827bd09bSSatish Balay /* determine the # of unique dof */ 599827bd09bSSatish Balay rvec_zero(lhs,m); 600827bd09bSSatish Balay rvec_set(lhs,1.0,n); 601827bd09bSSatish Balay gs_gop_hc(gs_handle,lhs,"+\0",level); 602827bd09bSSatish Balay rvec_zero(rsum,2); 603827bd09bSSatish Balay for (ct=i=0;i<n;i++) 604827bd09bSSatish Balay { 605827bd09bSSatish Balay if (lhs[i]!=0.0) 606827bd09bSSatish Balay {rsum[0]+=1.0/lhs[i]; rsum[1]+=lhs[i];} 607827bd09bSSatish Balay } 608827bd09bSSatish Balay grop_hc(rsum,rw,2,op,level); 609827bd09bSSatish Balay rsum[0]+=0.1; 610827bd09bSSatish Balay rsum[1]+=0.1; 611827bd09bSSatish Balay 612827bd09bSSatish Balay if (fabs(rsum[0]-rsum[1])>EPS) 613827bd09bSSatish Balay {shared=TRUE;} 614827bd09bSSatish Balay 61552f87cdaSBarry Smith xxt_handle->info->n_global=xxt_handle->info->m_global=(PetscInt) rsum[0]; 61652f87cdaSBarry Smith xxt_handle->mvi->n_global =xxt_handle->mvi->m_global =(PetscInt) rsum[0]; 617827bd09bSSatish Balay 618827bd09bSSatish Balay /* determine separator sets top down */ 619827bd09bSSatish Balay if (shared) 620827bd09bSSatish Balay { 621827bd09bSSatish Balay for (iptr=fo+n,id=my_id,mask=num_nodes>>1,edge=level;edge>0;edge--,mask>>=1) 622827bd09bSSatish Balay { 623827bd09bSSatish Balay /* set rsh of hc, fire, and collect lhs responses */ 624827bd09bSSatish Balay (id<mask) ? rvec_zero(lhs,m) : rvec_set(lhs,1.0,m); 625827bd09bSSatish Balay gs_gop_hc(gs_handle,lhs,"+\0",edge); 626827bd09bSSatish Balay 627827bd09bSSatish Balay /* set lsh of hc, fire, and collect rhs responses */ 628827bd09bSSatish Balay (id<mask) ? rvec_set(rhs,1.0,m) : rvec_zero(rhs,m); 629827bd09bSSatish Balay gs_gop_hc(gs_handle,rhs,"+\0",edge); 630827bd09bSSatish Balay 631827bd09bSSatish Balay for (i=0;i<n;i++) 632827bd09bSSatish Balay { 633827bd09bSSatish Balay if (id< mask) 634827bd09bSSatish Balay { 635827bd09bSSatish Balay if (lhs[i]!=0.0) 636827bd09bSSatish Balay {lhs[i]=1.0;} 637827bd09bSSatish Balay } 638827bd09bSSatish Balay if (id>=mask) 639827bd09bSSatish Balay { 640827bd09bSSatish Balay if (rhs[i]!=0.0) 641827bd09bSSatish Balay {rhs[i]=1.0;} 642827bd09bSSatish Balay } 643827bd09bSSatish Balay } 644827bd09bSSatish Balay 645827bd09bSSatish Balay if (id< mask) 646827bd09bSSatish Balay {gs_gop_hc(gs_handle,lhs,"+\0",edge-1);} 647827bd09bSSatish Balay else 648827bd09bSSatish Balay {gs_gop_hc(gs_handle,rhs,"+\0",edge-1);} 649827bd09bSSatish Balay 650827bd09bSSatish Balay /* count number of dofs I own that have signal and not in sep set */ 651827bd09bSSatish Balay rvec_zero(rsum,4); 652827bd09bSSatish Balay for (ivec_zero(sum,4),ct=i=0;i<n;i++) 653827bd09bSSatish Balay { 654827bd09bSSatish Balay if (!used[i]) 655827bd09bSSatish Balay { 656827bd09bSSatish Balay /* number of unmarked dofs on node */ 657827bd09bSSatish Balay ct++; 658827bd09bSSatish Balay /* number of dofs to be marked on lhs hc */ 659827bd09bSSatish Balay if (id< mask) 660827bd09bSSatish Balay { 661827bd09bSSatish Balay if (lhs[i]!=0.0) 662827bd09bSSatish Balay {sum[0]++; rsum[0]+=1.0/lhs[i];} 663827bd09bSSatish Balay } 664827bd09bSSatish Balay /* number of dofs to be marked on rhs hc */ 665827bd09bSSatish Balay if (id>=mask) 666827bd09bSSatish Balay { 667827bd09bSSatish Balay if (rhs[i]!=0.0) 668827bd09bSSatish Balay {sum[1]++; rsum[1]+=1.0/rhs[i];} 669827bd09bSSatish Balay } 670827bd09bSSatish Balay } 671827bd09bSSatish Balay } 672827bd09bSSatish Balay 673827bd09bSSatish Balay /* go for load balance - choose half with most unmarked dofs, bias LHS */ 674827bd09bSSatish Balay (id<mask) ? (sum[2]=ct) : (sum[3]=ct); 675827bd09bSSatish Balay (id<mask) ? (rsum[2]=ct) : (rsum[3]=ct); 676827bd09bSSatish Balay giop_hc(sum,w,4,op,edge); 677827bd09bSSatish Balay grop_hc(rsum,rw,4,op,edge); 678827bd09bSSatish Balay rsum[0]+=0.1; rsum[1]+=0.1; rsum[2]+=0.1; rsum[3]+=0.1; 679827bd09bSSatish Balay 680827bd09bSSatish Balay if (id<mask) 681827bd09bSSatish Balay { 682827bd09bSSatish Balay /* mark dofs I own that have signal and not in sep set */ 683827bd09bSSatish Balay for (ct=i=0;i<n;i++) 684827bd09bSSatish Balay { 685827bd09bSSatish Balay if ((!used[i])&&(lhs[i]!=0.0)) 686827bd09bSSatish Balay { 687827bd09bSSatish Balay ct++; nfo++; 688827bd09bSSatish Balay 689*c1235816SBarry Smith if (nfo>n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"nfo about to exceed n\n"); 690827bd09bSSatish Balay 691827bd09bSSatish Balay *--iptr = local2global[i]; 692827bd09bSSatish Balay used[i]=edge; 693827bd09bSSatish Balay } 694827bd09bSSatish Balay } 695827bd09bSSatish Balay if (ct>1) {ivec_sort(iptr,ct);} 696827bd09bSSatish Balay 697827bd09bSSatish Balay lnsep[edge]=ct; 69852f87cdaSBarry Smith nsep[edge]=(PetscInt) rsum[0]; 699827bd09bSSatish Balay dir [edge]=LEFT; 700827bd09bSSatish Balay } 701827bd09bSSatish Balay 702827bd09bSSatish Balay if (id>=mask) 703827bd09bSSatish Balay { 704827bd09bSSatish Balay /* mark dofs I own that have signal and not in sep set */ 705827bd09bSSatish Balay for (ct=i=0;i<n;i++) 706827bd09bSSatish Balay { 707827bd09bSSatish Balay if ((!used[i])&&(rhs[i]!=0.0)) 708827bd09bSSatish Balay { 709827bd09bSSatish Balay ct++; nfo++; 710827bd09bSSatish Balay 711*c1235816SBarry Smith if (nfo>n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"nfo about to exceed n\n"); 712827bd09bSSatish Balay 713827bd09bSSatish Balay *--iptr = local2global[i]; 714827bd09bSSatish Balay used[i]=edge; 715827bd09bSSatish Balay } 716827bd09bSSatish Balay } 717827bd09bSSatish Balay if (ct>1) {ivec_sort(iptr,ct);} 718827bd09bSSatish Balay 719827bd09bSSatish Balay lnsep[edge]=ct; 72052f87cdaSBarry Smith nsep[edge]= (PetscInt) rsum[1]; 721827bd09bSSatish Balay dir [edge]=RIGHT; 722827bd09bSSatish Balay } 723827bd09bSSatish Balay 724827bd09bSSatish Balay /* LATER or we can recur on these to order seps at this level */ 725827bd09bSSatish Balay /* do we need full set of separators for this? */ 726827bd09bSSatish Balay 727827bd09bSSatish Balay /* fold rhs hc into lower */ 728827bd09bSSatish Balay if (id>=mask) 729827bd09bSSatish Balay {id-=mask;} 730827bd09bSSatish Balay } 731827bd09bSSatish Balay } 732827bd09bSSatish Balay else 733827bd09bSSatish Balay { 734827bd09bSSatish Balay for (iptr=fo+n,id=my_id,mask=num_nodes>>1,edge=level;edge>0;edge--,mask>>=1) 735827bd09bSSatish Balay { 736827bd09bSSatish Balay /* set rsh of hc, fire, and collect lhs responses */ 737827bd09bSSatish Balay (id<mask) ? rvec_zero(lhs,m) : rvec_set(lhs,1.0,m); 738827bd09bSSatish Balay gs_gop_hc(gs_handle,lhs,"+\0",edge); 739827bd09bSSatish Balay 740827bd09bSSatish Balay /* set lsh of hc, fire, and collect rhs responses */ 741827bd09bSSatish Balay (id<mask) ? rvec_set(rhs,1.0,m) : rvec_zero(rhs,m); 742827bd09bSSatish Balay gs_gop_hc(gs_handle,rhs,"+\0",edge); 743827bd09bSSatish Balay 744827bd09bSSatish Balay /* count number of dofs I own that have signal and not in sep set */ 745827bd09bSSatish Balay for (ivec_zero(sum,4),ct=i=0;i<n;i++) 746827bd09bSSatish Balay { 747827bd09bSSatish Balay if (!used[i]) 748827bd09bSSatish Balay { 749827bd09bSSatish Balay /* number of unmarked dofs on node */ 750827bd09bSSatish Balay ct++; 751827bd09bSSatish Balay /* number of dofs to be marked on lhs hc */ 752827bd09bSSatish Balay if ((id< mask)&&(lhs[i]!=0.0)) {sum[0]++;} 753827bd09bSSatish Balay /* number of dofs to be marked on rhs hc */ 754827bd09bSSatish Balay if ((id>=mask)&&(rhs[i]!=0.0)) {sum[1]++;} 755827bd09bSSatish Balay } 756827bd09bSSatish Balay } 757827bd09bSSatish Balay 758827bd09bSSatish Balay /* go for load balance - choose half with most unmarked dofs, bias LHS */ 759827bd09bSSatish Balay (id<mask) ? (sum[2]=ct) : (sum[3]=ct); 760827bd09bSSatish Balay giop_hc(sum,w,4,op,edge); 761827bd09bSSatish Balay 762827bd09bSSatish Balay /* lhs hc wins */ 763827bd09bSSatish Balay if (sum[2]>=sum[3]) 764827bd09bSSatish Balay { 765827bd09bSSatish Balay if (id<mask) 766827bd09bSSatish Balay { 767827bd09bSSatish Balay /* mark dofs I own that have signal and not in sep set */ 768827bd09bSSatish Balay for (ct=i=0;i<n;i++) 769827bd09bSSatish Balay { 770827bd09bSSatish Balay if ((!used[i])&&(lhs[i]!=0.0)) 771827bd09bSSatish Balay { 772827bd09bSSatish Balay ct++; nfo++; 773827bd09bSSatish Balay *--iptr = local2global[i]; 774827bd09bSSatish Balay used[i]=edge; 775827bd09bSSatish Balay } 776827bd09bSSatish Balay } 777827bd09bSSatish Balay if (ct>1) {ivec_sort(iptr,ct);} 778827bd09bSSatish Balay lnsep[edge]=ct; 779827bd09bSSatish Balay } 780827bd09bSSatish Balay nsep[edge]=sum[0]; 781827bd09bSSatish Balay dir [edge]=LEFT; 782827bd09bSSatish Balay } 783827bd09bSSatish Balay /* rhs hc wins */ 784827bd09bSSatish Balay else 785827bd09bSSatish Balay { 786827bd09bSSatish Balay if (id>=mask) 787827bd09bSSatish Balay { 788827bd09bSSatish Balay /* mark dofs I own that have signal and not in sep set */ 789827bd09bSSatish Balay for (ct=i=0;i<n;i++) 790827bd09bSSatish Balay { 791827bd09bSSatish Balay if ((!used[i])&&(rhs[i]!=0.0)) 792827bd09bSSatish Balay { 793827bd09bSSatish Balay ct++; nfo++; 794827bd09bSSatish Balay *--iptr = local2global[i]; 795827bd09bSSatish Balay used[i]=edge; 796827bd09bSSatish Balay } 797827bd09bSSatish Balay } 798827bd09bSSatish Balay if (ct>1) {ivec_sort(iptr,ct);} 799827bd09bSSatish Balay lnsep[edge]=ct; 800827bd09bSSatish Balay } 801827bd09bSSatish Balay nsep[edge]=sum[1]; 802827bd09bSSatish Balay dir [edge]=RIGHT; 803827bd09bSSatish Balay } 804827bd09bSSatish Balay /* LATER or we can recur on these to order seps at this level */ 805827bd09bSSatish Balay /* do we need full set of separators for this? */ 806827bd09bSSatish Balay 807827bd09bSSatish Balay /* fold rhs hc into lower */ 808827bd09bSSatish Balay if (id>=mask) 809827bd09bSSatish Balay {id-=mask;} 810827bd09bSSatish Balay } 811827bd09bSSatish Balay } 812827bd09bSSatish Balay 813827bd09bSSatish Balay 814827bd09bSSatish Balay /* level 0 is on processor case - so mark the remainder */ 815827bd09bSSatish Balay for (ct=i=0;i<n;i++) 816827bd09bSSatish Balay { 817827bd09bSSatish Balay if (!used[i]) 818827bd09bSSatish Balay { 819827bd09bSSatish Balay ct++; nfo++; 820827bd09bSSatish Balay *--iptr = local2global[i]; 821827bd09bSSatish Balay used[i]=edge; 822827bd09bSSatish Balay } 823827bd09bSSatish Balay } 824827bd09bSSatish Balay if (ct>1) {ivec_sort(iptr,ct);} 825827bd09bSSatish Balay lnsep[edge]=ct; 826827bd09bSSatish Balay nsep [edge]=ct; 827827bd09bSSatish Balay dir [edge]=LEFT; 828827bd09bSSatish Balay 829827bd09bSSatish Balay xxt_handle->info->nsep=nsep; 830827bd09bSSatish Balay xxt_handle->info->lnsep=lnsep; 831827bd09bSSatish Balay xxt_handle->info->fo=fo; 832827bd09bSSatish Balay xxt_handle->info->nfo=nfo; 833827bd09bSSatish Balay 834a501084fSBarry Smith free(dir); 835a501084fSBarry Smith free(lhs); 836a501084fSBarry Smith free(rhs); 837a501084fSBarry Smith free(used); 8380924e98cSBarry Smith PetscFunctionReturn(0); 839827bd09bSSatish Balay } 840827bd09bSSatish Balay 8417b1ae94cSBarry Smith /**************************************xxt.c***********************************/ 84252f87cdaSBarry Smith static mv_info *set_mvi(PetscInt *local2global, PetscInt n, PetscInt m, void *matvec, void *grid_data) 843827bd09bSSatish Balay { 844827bd09bSSatish Balay mv_info *mvi; 845827bd09bSSatish Balay 846827bd09bSSatish Balay 847a501084fSBarry Smith mvi = (mv_info*)malloc(sizeof(mv_info)); 848827bd09bSSatish Balay mvi->n=n; 849827bd09bSSatish Balay mvi->m=m; 850827bd09bSSatish Balay mvi->n_global=-1; 851827bd09bSSatish Balay mvi->m_global=-1; 85252f87cdaSBarry Smith mvi->local2global=(PetscInt*)malloc((m+1)*sizeof(PetscInt)); 853827bd09bSSatish Balay ivec_copy(mvi->local2global,local2global,m); 854827bd09bSSatish Balay mvi->local2global[m] = INT_MAX; 855a501084fSBarry Smith mvi->matvec=(PetscErrorCode (*)(mv_info*,PetscScalar*,PetscScalar*))matvec; 856827bd09bSSatish Balay mvi->grid_data=grid_data; 857827bd09bSSatish Balay 858827bd09bSSatish Balay /* set xxt communication handle to perform restricted matvec */ 859827bd09bSSatish Balay mvi->gs_handle = gs_init(local2global, m, num_nodes); 860827bd09bSSatish Balay 861827bd09bSSatish Balay return(mvi); 862827bd09bSSatish Balay } 863827bd09bSSatish Balay 8647b1ae94cSBarry Smith /**************************************xxt.c***********************************/ 8650924e98cSBarry Smith static PetscErrorCode do_matvec(mv_info *A, PetscScalar *v, PetscScalar *u) 866827bd09bSSatish Balay { 8670924e98cSBarry Smith PetscFunctionBegin; 868827bd09bSSatish Balay A->matvec((mv_info*)A->grid_data,v,u); 8690924e98cSBarry Smith PetscFunctionReturn(0); 870827bd09bSSatish Balay } 871827bd09bSSatish Balay 872827bd09bSSatish Balay 873827bd09bSSatish Balay 874