1dba47a55SKris Buschelman #define PETSCKSP_DLL 2827bd09bSSatish Balay 3827bd09bSSatish Balay /***********************************gs.c*************************************** 4827bd09bSSatish Balay 5827bd09bSSatish Balay Author: Henry M. Tufo III 6827bd09bSSatish Balay 7827bd09bSSatish Balay e-mail: hmt@cs.brown.edu 8827bd09bSSatish Balay 9827bd09bSSatish Balay snail-mail: 10827bd09bSSatish Balay Division of Applied Mathematics 11827bd09bSSatish Balay Brown University 12827bd09bSSatish Balay Providence, RI 02912 13827bd09bSSatish Balay 14827bd09bSSatish Balay Last Modification: 15827bd09bSSatish Balay 6.21.97 16827bd09bSSatish Balay ************************************gs.c**************************************/ 17827bd09bSSatish Balay 18827bd09bSSatish Balay /***********************************gs.c*************************************** 19827bd09bSSatish Balay File Description: 20827bd09bSSatish Balay ----------------- 21827bd09bSSatish Balay 22827bd09bSSatish Balay ************************************gs.c**************************************/ 23827bd09bSSatish Balay 247758a8cdSBarry Smith #include "src/ksp/pc/impls/tfs/tfs.h" 2539945688SSatish Balay 26827bd09bSSatish Balay /* default length of number of items via tree - doubles if exceeded */ 27827bd09bSSatish Balay #define TREE_BUF_SZ 2048; 28827bd09bSSatish Balay #define GS_VEC_SZ 1 29827bd09bSSatish Balay 30827bd09bSSatish Balay 31827bd09bSSatish Balay 32827bd09bSSatish Balay /***********************************gs.c*************************************** 33827bd09bSSatish Balay Type: struct gather_scatter_id 34827bd09bSSatish Balay ------------------------------ 35827bd09bSSatish Balay 36827bd09bSSatish Balay ************************************gs.c**************************************/ 37827bd09bSSatish Balay typedef struct gather_scatter_id { 3852f87cdaSBarry Smith PetscInt id; 3952f87cdaSBarry Smith PetscInt nel_min; 4052f87cdaSBarry Smith PetscInt nel_max; 4152f87cdaSBarry Smith PetscInt nel_sum; 4252f87cdaSBarry Smith PetscInt negl; 4352f87cdaSBarry Smith PetscInt gl_max; 4452f87cdaSBarry Smith PetscInt gl_min; 4552f87cdaSBarry Smith PetscInt repeats; 4652f87cdaSBarry Smith PetscInt ordered; 4752f87cdaSBarry Smith PetscInt positive; 48a501084fSBarry Smith PetscScalar *vals; 49827bd09bSSatish Balay 50827bd09bSSatish Balay /* bit mask info */ 5152f87cdaSBarry Smith PetscInt *my_proc_mask; 5252f87cdaSBarry Smith PetscInt mask_sz; 5352f87cdaSBarry Smith PetscInt *ngh_buf; 5452f87cdaSBarry Smith PetscInt ngh_buf_sz; 5552f87cdaSBarry Smith PetscInt *nghs; 5652f87cdaSBarry Smith PetscInt num_nghs; 5752f87cdaSBarry Smith PetscInt max_nghs; 5852f87cdaSBarry Smith PetscInt *pw_nghs; 5952f87cdaSBarry Smith PetscInt num_pw_nghs; 6052f87cdaSBarry Smith PetscInt *tree_nghs; 6152f87cdaSBarry Smith PetscInt num_tree_nghs; 62827bd09bSSatish Balay 6352f87cdaSBarry Smith PetscInt num_loads; 64827bd09bSSatish Balay 65827bd09bSSatish Balay /* repeats == true -> local info */ 6652f87cdaSBarry Smith PetscInt nel; /* number of unique elememts */ 6752f87cdaSBarry Smith PetscInt *elms; /* of size nel */ 6852f87cdaSBarry Smith PetscInt nel_total; 6952f87cdaSBarry Smith PetscInt *local_elms; /* of size nel_total */ 7052f87cdaSBarry Smith PetscInt *companion; /* of size nel_total */ 71827bd09bSSatish Balay 72827bd09bSSatish Balay /* local info */ 7352f87cdaSBarry Smith PetscInt num_local_total; 7452f87cdaSBarry Smith PetscInt local_strength; 7552f87cdaSBarry Smith PetscInt num_local; 7652f87cdaSBarry Smith PetscInt *num_local_reduce; 7752f87cdaSBarry Smith PetscInt **local_reduce; 7852f87cdaSBarry Smith PetscInt num_local_gop; 7952f87cdaSBarry Smith PetscInt *num_gop_local_reduce; 8052f87cdaSBarry Smith PetscInt **gop_local_reduce; 81827bd09bSSatish Balay 82827bd09bSSatish Balay /* pairwise info */ 8352f87cdaSBarry Smith PetscInt level; 8452f87cdaSBarry Smith PetscInt num_pairs; 8552f87cdaSBarry Smith PetscInt max_pairs; 8652f87cdaSBarry Smith PetscInt loc_node_pairs; 8752f87cdaSBarry Smith PetscInt max_node_pairs; 8852f87cdaSBarry Smith PetscInt min_node_pairs; 8952f87cdaSBarry Smith PetscInt avg_node_pairs; 9052f87cdaSBarry Smith PetscInt *pair_list; 9152f87cdaSBarry Smith PetscInt *msg_sizes; 9252f87cdaSBarry Smith PetscInt **node_list; 9352f87cdaSBarry Smith PetscInt len_pw_list; 9452f87cdaSBarry Smith PetscInt *pw_elm_list; 95a501084fSBarry Smith PetscScalar *pw_vals; 96827bd09bSSatish Balay 97827bd09bSSatish Balay MPI_Request *msg_ids_in; 98827bd09bSSatish Balay MPI_Request *msg_ids_out; 99827bd09bSSatish Balay 100a501084fSBarry Smith PetscScalar *out; 101a501084fSBarry Smith PetscScalar *in; 10252f87cdaSBarry Smith PetscInt msg_total; 103827bd09bSSatish Balay 104827bd09bSSatish Balay /* tree - crystal accumulator info */ 10552f87cdaSBarry Smith PetscInt max_left_over; 10652f87cdaSBarry Smith PetscInt *pre; 10752f87cdaSBarry Smith PetscInt *in_num; 10852f87cdaSBarry Smith PetscInt *out_num; 10952f87cdaSBarry Smith PetscInt **in_list; 11052f87cdaSBarry Smith PetscInt **out_list; 111827bd09bSSatish Balay 112827bd09bSSatish Balay /* new tree work*/ 11352f87cdaSBarry Smith PetscInt tree_nel; 11452f87cdaSBarry Smith PetscInt *tree_elms; 115a501084fSBarry Smith PetscScalar *tree_buf; 116a501084fSBarry Smith PetscScalar *tree_work; 117827bd09bSSatish Balay 11852f87cdaSBarry Smith PetscInt tree_map_sz; 11952f87cdaSBarry Smith PetscInt *tree_map_in; 12052f87cdaSBarry Smith PetscInt *tree_map_out; 121827bd09bSSatish Balay 122827bd09bSSatish Balay /* current memory status */ 12352f87cdaSBarry Smith PetscInt gl_bss_min; 12452f87cdaSBarry Smith PetscInt gl_perm_min; 125827bd09bSSatish Balay 126827bd09bSSatish Balay /* max segment size for gs_gop_vec() */ 12752f87cdaSBarry Smith PetscInt vec_sz; 128827bd09bSSatish Balay 129827bd09bSSatish Balay /* hack to make paul happy */ 130827bd09bSSatish Balay MPI_Comm gs_comm; 131827bd09bSSatish Balay 132827bd09bSSatish Balay } gs_id; 133827bd09bSSatish Balay 13452f87cdaSBarry Smith static gs_id *gsi_check_args(PetscInt *elms, PetscInt nel, PetscInt level); 1353fdc5746SBarry Smith static PetscErrorCode gsi_via_bit_mask(gs_id *gs); 1363fdc5746SBarry Smith static PetscErrorCode get_ngh_buf(gs_id *gs); 1373fdc5746SBarry Smith static PetscErrorCode set_pairwise(gs_id *gs); 138827bd09bSSatish Balay static gs_id * gsi_new(void); 1393fdc5746SBarry Smith static PetscErrorCode set_tree(gs_id *gs); 140827bd09bSSatish Balay 141827bd09bSSatish Balay /* same for all but vector flavor */ 1423fdc5746SBarry Smith static PetscErrorCode gs_gop_local_out(gs_id *gs, PetscScalar *vals); 143827bd09bSSatish Balay /* vector flavor */ 14452f87cdaSBarry Smith static PetscErrorCode gs_gop_vec_local_out(gs_id *gs, PetscScalar *vals, PetscInt step); 145827bd09bSSatish Balay 14652f87cdaSBarry Smith static PetscErrorCode gs_gop_vec_plus(gs_id *gs, PetscScalar *in_vals, PetscInt step); 14752f87cdaSBarry Smith static PetscErrorCode gs_gop_vec_pairwise_plus(gs_id *gs, PetscScalar *in_vals, PetscInt step); 14852f87cdaSBarry Smith static PetscErrorCode gs_gop_vec_local_plus(gs_id *gs, PetscScalar *vals, PetscInt step); 14952f87cdaSBarry Smith static PetscErrorCode gs_gop_vec_local_in_plus(gs_id *gs, PetscScalar *vals, PetscInt step); 15052f87cdaSBarry Smith static PetscErrorCode gs_gop_vec_tree_plus(gs_id *gs, PetscScalar *vals, PetscInt step); 151827bd09bSSatish Balay 152827bd09bSSatish Balay 1533fdc5746SBarry Smith static PetscErrorCode gs_gop_plus(gs_id *gs, PetscScalar *in_vals); 1543fdc5746SBarry Smith static PetscErrorCode gs_gop_pairwise_plus(gs_id *gs, PetscScalar *in_vals); 1553fdc5746SBarry Smith static PetscErrorCode gs_gop_local_plus(gs_id *gs, PetscScalar *vals); 1563fdc5746SBarry Smith static PetscErrorCode gs_gop_local_in_plus(gs_id *gs, PetscScalar *vals); 1573fdc5746SBarry Smith static PetscErrorCode gs_gop_tree_plus(gs_id *gs, PetscScalar *vals); 158827bd09bSSatish Balay 15952f87cdaSBarry Smith static PetscErrorCode gs_gop_plus_hc(gs_id *gs, PetscScalar *in_vals, PetscInt dim); 16052f87cdaSBarry Smith static PetscErrorCode gs_gop_pairwise_plus_hc(gs_id *gs, PetscScalar *in_vals, PetscInt dim); 16152f87cdaSBarry Smith static PetscErrorCode gs_gop_tree_plus_hc(gs_id *gs, PetscScalar *vals, PetscInt dim); 162827bd09bSSatish Balay 1633fdc5746SBarry Smith static PetscErrorCode gs_gop_times(gs_id *gs, PetscScalar *in_vals); 1643fdc5746SBarry Smith static PetscErrorCode gs_gop_pairwise_times(gs_id *gs, PetscScalar *in_vals); 1653fdc5746SBarry Smith static PetscErrorCode gs_gop_local_times(gs_id *gs, PetscScalar *vals); 1663fdc5746SBarry Smith static PetscErrorCode gs_gop_local_in_times(gs_id *gs, PetscScalar *vals); 1673fdc5746SBarry Smith static PetscErrorCode gs_gop_tree_times(gs_id *gs, PetscScalar *vals); 168827bd09bSSatish Balay 1693fdc5746SBarry Smith static PetscErrorCode gs_gop_min(gs_id *gs, PetscScalar *in_vals); 1703fdc5746SBarry Smith static PetscErrorCode gs_gop_pairwise_min(gs_id *gs, PetscScalar *in_vals); 1713fdc5746SBarry Smith static PetscErrorCode gs_gop_local_min(gs_id *gs, PetscScalar *vals); 1723fdc5746SBarry Smith static PetscErrorCode gs_gop_local_in_min(gs_id *gs, PetscScalar *vals); 1733fdc5746SBarry Smith static PetscErrorCode gs_gop_tree_min(gs_id *gs, PetscScalar *vals); 174827bd09bSSatish Balay 1753fdc5746SBarry Smith static PetscErrorCode gs_gop_min_abs(gs_id *gs, PetscScalar *in_vals); 1763fdc5746SBarry Smith static PetscErrorCode gs_gop_pairwise_min_abs(gs_id *gs, PetscScalar *in_vals); 1773fdc5746SBarry Smith static PetscErrorCode gs_gop_local_min_abs(gs_id *gs, PetscScalar *vals); 1783fdc5746SBarry Smith static PetscErrorCode gs_gop_local_in_min_abs(gs_id *gs, PetscScalar *vals); 1793fdc5746SBarry Smith static PetscErrorCode gs_gop_tree_min_abs(gs_id *gs, PetscScalar *vals); 180827bd09bSSatish Balay 1813fdc5746SBarry Smith static PetscErrorCode gs_gop_max(gs_id *gs, PetscScalar *in_vals); 1823fdc5746SBarry Smith static PetscErrorCode gs_gop_pairwise_max(gs_id *gs, PetscScalar *in_vals); 1833fdc5746SBarry Smith static PetscErrorCode gs_gop_local_max(gs_id *gs, PetscScalar *vals); 1843fdc5746SBarry Smith static PetscErrorCode gs_gop_local_in_max(gs_id *gs, PetscScalar *vals); 1853fdc5746SBarry Smith static PetscErrorCode gs_gop_tree_max(gs_id *gs, PetscScalar *vals); 186827bd09bSSatish Balay 1873fdc5746SBarry Smith static PetscErrorCode gs_gop_max_abs(gs_id *gs, PetscScalar *in_vals); 1883fdc5746SBarry Smith static PetscErrorCode gs_gop_pairwise_max_abs(gs_id *gs, PetscScalar *in_vals); 1893fdc5746SBarry Smith static PetscErrorCode gs_gop_local_max_abs(gs_id *gs, PetscScalar *vals); 1903fdc5746SBarry Smith static PetscErrorCode gs_gop_local_in_max_abs(gs_id *gs, PetscScalar *vals); 1913fdc5746SBarry Smith static PetscErrorCode gs_gop_tree_max_abs(gs_id *gs, PetscScalar *vals); 192827bd09bSSatish Balay 1933fdc5746SBarry Smith static PetscErrorCode gs_gop_exists(gs_id *gs, PetscScalar *in_vals); 1943fdc5746SBarry Smith static PetscErrorCode gs_gop_pairwise_exists(gs_id *gs, PetscScalar *in_vals); 1953fdc5746SBarry Smith static PetscErrorCode gs_gop_local_exists(gs_id *gs, PetscScalar *vals); 1963fdc5746SBarry Smith static PetscErrorCode gs_gop_local_in_exists(gs_id *gs, PetscScalar *vals); 1973fdc5746SBarry Smith static PetscErrorCode gs_gop_tree_exists(gs_id *gs, PetscScalar *vals); 198827bd09bSSatish Balay 1993fdc5746SBarry Smith static PetscErrorCode gs_gop_pairwise_binary(gs_id *gs, PetscScalar *in_vals, rbfp fct); 2003fdc5746SBarry Smith static PetscErrorCode gs_gop_local_binary(gs_id *gs, PetscScalar *vals, rbfp fct); 2013fdc5746SBarry Smith static PetscErrorCode gs_gop_local_in_binary(gs_id *gs, PetscScalar *vals, rbfp fct); 2023fdc5746SBarry Smith static PetscErrorCode gs_gop_tree_binary(gs_id *gs, PetscScalar *vals, rbfp fct); 203827bd09bSSatish Balay 204827bd09bSSatish Balay 205827bd09bSSatish Balay 206827bd09bSSatish Balay /* global vars */ 207827bd09bSSatish Balay /* from comm.c module */ 208827bd09bSSatish Balay 20952f87cdaSBarry Smith static PetscInt num_gs_ids = 0; 210827bd09bSSatish Balay 211827bd09bSSatish Balay /* should make this dynamic ... later */ 21252f87cdaSBarry Smith static PetscInt msg_buf=MAX_MSG_BUF; 21352f87cdaSBarry Smith static PetscInt vec_sz=GS_VEC_SZ; 21452f87cdaSBarry Smith static PetscInt *tree_buf=NULL; 21552f87cdaSBarry Smith static PetscInt tree_buf_sz=0; 21652f87cdaSBarry Smith static PetscInt ntree=0; 217827bd09bSSatish Balay 218827bd09bSSatish Balay 219827bd09bSSatish Balay /****************************************************************************** 220827bd09bSSatish Balay Function: gs_init_() 221827bd09bSSatish Balay 222827bd09bSSatish Balay Input : 223827bd09bSSatish Balay Output: 224827bd09bSSatish Balay Return: 225827bd09bSSatish Balay Description: 226827bd09bSSatish Balay ******************************************************************************/ 22752f87cdaSBarry Smith PetscErrorCode gs_init_vec_sz(PetscInt size) 228827bd09bSSatish Balay { 2293fdc5746SBarry Smith PetscFunctionBegin; 230827bd09bSSatish Balay vec_sz = size; 2313fdc5746SBarry Smith PetscFunctionReturn(0); 232827bd09bSSatish Balay } 233827bd09bSSatish Balay 234827bd09bSSatish Balay /****************************************************************************** 235827bd09bSSatish Balay Function: gs_init_() 236827bd09bSSatish Balay 237827bd09bSSatish Balay Input : 238827bd09bSSatish Balay Output: 239827bd09bSSatish Balay Return: 240827bd09bSSatish Balay Description: 241827bd09bSSatish Balay ******************************************************************************/ 24252f87cdaSBarry Smith PetscErrorCode gs_init_msg_buf_sz(PetscInt buf_size) 243827bd09bSSatish Balay { 2443fdc5746SBarry Smith PetscFunctionBegin; 245827bd09bSSatish Balay msg_buf = buf_size; 2463fdc5746SBarry Smith PetscFunctionReturn(0); 247827bd09bSSatish Balay } 248827bd09bSSatish Balay 249827bd09bSSatish Balay /****************************************************************************** 250827bd09bSSatish Balay Function: gs_init() 251827bd09bSSatish Balay 252827bd09bSSatish Balay Input : 253827bd09bSSatish Balay 254827bd09bSSatish Balay Output: 255827bd09bSSatish Balay 256827bd09bSSatish Balay RETURN: 257827bd09bSSatish Balay 258827bd09bSSatish Balay Description: 259827bd09bSSatish Balay ******************************************************************************/ 26052f87cdaSBarry Smith gs_id *gs_init( PetscInt *elms, PetscInt nel, PetscInt level) 261827bd09bSSatish Balay { 262a501084fSBarry Smith gs_id *gs; 263827bd09bSSatish Balay MPI_Group gs_group; 264827bd09bSSatish Balay MPI_Comm gs_comm; 265827bd09bSSatish Balay 2663fdc5746SBarry Smith PetscFunctionBegin; 267827bd09bSSatish Balay /* ensure that communication package has been initialized */ 268827bd09bSSatish Balay comm_init(); 269827bd09bSSatish Balay 270827bd09bSSatish Balay 271827bd09bSSatish Balay /* determines if we have enough dynamic/semi-static memory */ 272827bd09bSSatish Balay /* checks input, allocs and sets gd_id template */ 273827bd09bSSatish Balay gs = gsi_check_args(elms,nel,level); 274827bd09bSSatish Balay 275827bd09bSSatish Balay /* only bit mask version up and working for the moment */ 276827bd09bSSatish Balay /* LATER :: get int list version working for sparse pblms */ 277827bd09bSSatish Balay gsi_via_bit_mask(gs); 278827bd09bSSatish Balay 279827bd09bSSatish Balay 280827bd09bSSatish Balay MPI_Comm_group(MPI_COMM_WORLD,&gs_group); 281827bd09bSSatish Balay MPI_Comm_create(MPI_COMM_WORLD,gs_group,&gs_comm); 282827bd09bSSatish Balay gs->gs_comm=gs_comm; 283827bd09bSSatish Balay 284827bd09bSSatish Balay return(gs); 285827bd09bSSatish Balay } 286827bd09bSSatish Balay 287827bd09bSSatish Balay 288827bd09bSSatish Balay 289827bd09bSSatish Balay /****************************************************************************** 290827bd09bSSatish Balay Function: gsi_new() 291827bd09bSSatish Balay 292827bd09bSSatish Balay Input : 293827bd09bSSatish Balay Output: 294827bd09bSSatish Balay Return: 295827bd09bSSatish Balay Description: 296827bd09bSSatish Balay 297827bd09bSSatish Balay elm list must >= 0!!! 298827bd09bSSatish Balay elm repeats allowed 299827bd09bSSatish Balay ******************************************************************************/ 3000924e98cSBarry Smith static gs_id *gsi_new(void) 301827bd09bSSatish Balay { 302827bd09bSSatish Balay gs_id *gs; 303330ea6edSBarry Smith gs = (gs_id *) malloc(sizeof(gs_id)); 304330ea6edSBarry Smith PetscMemzero(gs,sizeof(gs_id)); 305827bd09bSSatish Balay return(gs); 306827bd09bSSatish Balay } 307827bd09bSSatish Balay 308827bd09bSSatish Balay 309827bd09bSSatish Balay 310827bd09bSSatish Balay /****************************************************************************** 311827bd09bSSatish Balay Function: gsi_check_args() 312827bd09bSSatish Balay 313827bd09bSSatish Balay Input : 314827bd09bSSatish Balay Output: 315827bd09bSSatish Balay Return: 316827bd09bSSatish Balay Description: 317827bd09bSSatish Balay 318827bd09bSSatish Balay elm list must >= 0!!! 319827bd09bSSatish Balay elm repeats allowed 320827bd09bSSatish Balay local working copy of elms is sorted 321827bd09bSSatish Balay ******************************************************************************/ 32252f87cdaSBarry Smith static gs_id * gsi_check_args(PetscInt *in_elms, PetscInt nel, PetscInt level) 323827bd09bSSatish Balay { 32452f87cdaSBarry Smith PetscInt i, j, k, t2; 32552f87cdaSBarry Smith PetscInt *companion, *elms, *unique, *iptr; 32652f87cdaSBarry Smith PetscInt num_local=0, *num_to_reduce, **local_reduce; 32752f87cdaSBarry Smith PetscInt oprs[] = {NON_UNIFORM,GL_MIN,GL_MAX,GL_ADD,GL_MIN,GL_MAX,GL_MIN,GL_B_AND}; 32852f87cdaSBarry Smith PetscInt vals[sizeof(oprs)/sizeof(oprs[0])-1]; 32952f87cdaSBarry Smith PetscInt work[sizeof(oprs)/sizeof(oprs[0])-1]; 330827bd09bSSatish Balay gs_id *gs; 331*d1528f56SBarry Smith PetscErrorCode ierr; 332827bd09bSSatish Balay 333827bd09bSSatish Balay 334827bd09bSSatish Balay if (!in_elms) 335388eb383SBarry Smith {SETERRABORT(PETSC_COMM_WORLD,PETSC_ERR_PLIB,"elms point to nothing!!!\n");} 336827bd09bSSatish Balay 337827bd09bSSatish Balay if (nel<0) 338388eb383SBarry Smith {SETERRABORT(PETSC_COMM_WORLD,PETSC_ERR_PLIB,"can't have fewer than 0 elms!!!\n");} 339827bd09bSSatish Balay 340827bd09bSSatish Balay if (nel==0) 341*d1528f56SBarry Smith {ierr = PetscInfo(0,"I don't have any elements!!!\n");} 342827bd09bSSatish Balay 343827bd09bSSatish Balay /* get space for gs template */ 344827bd09bSSatish Balay gs = gsi_new(); 345827bd09bSSatish Balay gs->id = ++num_gs_ids; 346827bd09bSSatish Balay 347827bd09bSSatish Balay /* hmt 6.4.99 */ 348827bd09bSSatish Balay /* caller can set global ids that don't participate to 0 */ 349827bd09bSSatish Balay /* gs_init ignores all zeros in elm list */ 350827bd09bSSatish Balay /* negative global ids are still invalid */ 351827bd09bSSatish Balay for (i=j=0;i<nel;i++) 352827bd09bSSatish Balay {if (in_elms[i]!=0) {j++;}} 353827bd09bSSatish Balay 354827bd09bSSatish Balay k=nel; nel=j; 355827bd09bSSatish Balay 356827bd09bSSatish Balay /* copy over in_elms list and create inverse map */ 35752f87cdaSBarry Smith elms = (PetscInt*) malloc((nel+1)*sizeof(PetscInt)); 35852f87cdaSBarry Smith companion = (PetscInt*) malloc(nel*sizeof(PetscInt)); 3591d7d0905SBarry Smith 360827bd09bSSatish Balay for (i=j=0;i<k;i++) 361827bd09bSSatish Balay { 362827bd09bSSatish Balay if (in_elms[i]!=0) 363827bd09bSSatish Balay {elms[j] = in_elms[i]; companion[j++] = i;} 364827bd09bSSatish Balay } 365827bd09bSSatish Balay 366827bd09bSSatish Balay if (j!=nel) 367388eb383SBarry Smith {SETERRABORT(PETSC_COMM_WORLD,PETSC_ERR_PLIB,"nel j mismatch!\n");} 368827bd09bSSatish Balay 369827bd09bSSatish Balay /* pre-pass ... check to see if sorted */ 370827bd09bSSatish Balay elms[nel] = INT_MAX; 371827bd09bSSatish Balay iptr = elms; 372827bd09bSSatish Balay unique = elms+1; 373827bd09bSSatish Balay j=0; 374827bd09bSSatish Balay while (*iptr!=INT_MAX) 375827bd09bSSatish Balay { 376827bd09bSSatish Balay if (*iptr++>*unique++) 377827bd09bSSatish Balay {j=1; break;} 378827bd09bSSatish Balay } 379827bd09bSSatish Balay 380827bd09bSSatish Balay /* set up inverse map */ 381827bd09bSSatish Balay if (j) 382827bd09bSSatish Balay { 383*d1528f56SBarry Smith ierr = PetscInfo(0,"gsi_check_args() :: elm list *not* sorted!\n"); 384827bd09bSSatish Balay SMI_sort((void*)elms, (void*)companion, nel, SORT_INTEGER); 385827bd09bSSatish Balay } 386827bd09bSSatish Balay else 387*d1528f56SBarry Smith {ierr = PetscInfo(0,"gsi_check_args() :: elm list sorted!\n");} 388827bd09bSSatish Balay elms[nel] = INT_MIN; 389827bd09bSSatish Balay 390827bd09bSSatish Balay /* first pass */ 391827bd09bSSatish Balay /* determine number of unique elements, check pd */ 392827bd09bSSatish Balay for (i=k=0;i<nel;i+=j) 393827bd09bSSatish Balay { 394827bd09bSSatish Balay t2 = elms[i]; 395827bd09bSSatish Balay j=++i; 396827bd09bSSatish Balay 397827bd09bSSatish Balay /* clump 'em for now */ 398827bd09bSSatish Balay while (elms[j]==t2) {j++;} 399827bd09bSSatish Balay 400827bd09bSSatish Balay /* how many together and num local */ 401827bd09bSSatish Balay if (j-=i) 402827bd09bSSatish Balay {num_local++; k+=j;} 403827bd09bSSatish Balay } 404827bd09bSSatish Balay 405827bd09bSSatish Balay /* how many unique elements? */ 406827bd09bSSatish Balay gs->repeats=k; 407827bd09bSSatish Balay gs->nel = nel-k; 408827bd09bSSatish Balay 409827bd09bSSatish Balay 410827bd09bSSatish Balay /* number of repeats? */ 411827bd09bSSatish Balay gs->num_local = num_local; 412827bd09bSSatish Balay num_local+=2; 41352f87cdaSBarry Smith gs->local_reduce=local_reduce=(PetscInt **)malloc(num_local*sizeof(PetscInt*)); 41452f87cdaSBarry Smith gs->num_local_reduce=num_to_reduce=(PetscInt*) malloc(num_local*sizeof(PetscInt)); 415827bd09bSSatish Balay 41652f87cdaSBarry Smith unique = (PetscInt*) malloc((gs->nel+1)*sizeof(PetscInt)); 417827bd09bSSatish Balay gs->elms = unique; 418827bd09bSSatish Balay gs->nel_total = nel; 419827bd09bSSatish Balay gs->local_elms = elms; 420827bd09bSSatish Balay gs->companion = companion; 421827bd09bSSatish Balay 422827bd09bSSatish Balay /* compess map as well as keep track of local ops */ 423827bd09bSSatish Balay for (num_local=i=j=0;i<gs->nel;i++) 424827bd09bSSatish Balay { 425827bd09bSSatish Balay k=j; 426827bd09bSSatish Balay t2 = unique[i] = elms[j]; 427827bd09bSSatish Balay companion[i] = companion[j]; 428827bd09bSSatish Balay 429827bd09bSSatish Balay while (elms[j]==t2) {j++;} 430827bd09bSSatish Balay 431827bd09bSSatish Balay if ((t2=(j-k))>1) 432827bd09bSSatish Balay { 433827bd09bSSatish Balay /* number together */ 434827bd09bSSatish Balay num_to_reduce[num_local] = t2++; 43552f87cdaSBarry Smith iptr = local_reduce[num_local++] = (PetscInt*)malloc(t2*sizeof(PetscInt)); 436827bd09bSSatish Balay 437827bd09bSSatish Balay /* to use binary searching don't remap until we check intersection */ 438827bd09bSSatish Balay *iptr++ = i; 439827bd09bSSatish Balay 440827bd09bSSatish Balay /* note that we're skipping the first one */ 441827bd09bSSatish Balay while (++k<j) 442827bd09bSSatish Balay {*(iptr++) = companion[k];} 443827bd09bSSatish Balay *iptr = -1; 444827bd09bSSatish Balay } 445827bd09bSSatish Balay } 446827bd09bSSatish Balay 447827bd09bSSatish Balay /* sentinel for ngh_buf */ 448827bd09bSSatish Balay unique[gs->nel]=INT_MAX; 449827bd09bSSatish Balay 450827bd09bSSatish Balay /* for two partition sort hack */ 451827bd09bSSatish Balay num_to_reduce[num_local] = 0; 452827bd09bSSatish Balay local_reduce[num_local] = NULL; 453827bd09bSSatish Balay num_to_reduce[++num_local] = 0; 454827bd09bSSatish Balay local_reduce[num_local] = NULL; 455827bd09bSSatish Balay 456827bd09bSSatish Balay /* load 'em up */ 457827bd09bSSatish Balay /* note one extra to hold NON_UNIFORM flag!!! */ 458827bd09bSSatish Balay vals[2] = vals[1] = vals[0] = nel; 459827bd09bSSatish Balay if (gs->nel>0) 460827bd09bSSatish Balay { 4611d7d0905SBarry Smith vals[3] = unique[0]; 4621d7d0905SBarry Smith vals[4] = unique[gs->nel-1]; 463827bd09bSSatish Balay } 464827bd09bSSatish Balay else 465827bd09bSSatish Balay { 4661d7d0905SBarry Smith vals[3] = INT_MAX; 4671d7d0905SBarry Smith vals[4] = INT_MIN; 468827bd09bSSatish Balay } 469827bd09bSSatish Balay vals[5] = level; 470827bd09bSSatish Balay vals[6] = num_gs_ids; 471827bd09bSSatish Balay 472827bd09bSSatish Balay /* GLOBAL: send 'em out */ 473827bd09bSSatish Balay giop(vals,work,sizeof(oprs)/sizeof(oprs[0])-1,oprs); 474827bd09bSSatish Balay 475827bd09bSSatish Balay /* must be semi-pos def - only pairwise depends on this */ 476827bd09bSSatish Balay /* LATER - remove this restriction */ 477827bd09bSSatish Balay if (vals[3]<0) 478388eb383SBarry Smith {SETERRABORT(PETSC_COMM_WORLD,PETSC_ERR_PLIB,"gsi_check_args() :: system not semi-pos def \n");} 479827bd09bSSatish Balay 480827bd09bSSatish Balay if (vals[4]==INT_MAX) 481388eb383SBarry Smith {SETERRABORT(PETSC_COMM_WORLD,PETSC_ERR_PLIB,"gsi_check_args() :: system ub too large !\n");} 482827bd09bSSatish Balay 483827bd09bSSatish Balay gs->nel_min = vals[0]; 484827bd09bSSatish Balay gs->nel_max = vals[1]; 485827bd09bSSatish Balay gs->nel_sum = vals[2]; 486827bd09bSSatish Balay gs->gl_min = vals[3]; 487827bd09bSSatish Balay gs->gl_max = vals[4]; 488827bd09bSSatish Balay gs->negl = vals[4]-vals[3]+1; 489827bd09bSSatish Balay 490827bd09bSSatish Balay if (gs->negl<=0) 491388eb383SBarry Smith {SETERRABORT(PETSC_COMM_WORLD,PETSC_ERR_PLIB,"gsi_check_args() :: system empty or neg :: %d\n");} 492827bd09bSSatish Balay 493827bd09bSSatish Balay /* LATER :: add level == -1 -> program selects level */ 494827bd09bSSatish Balay if (vals[5]<0) 495827bd09bSSatish Balay {vals[5]=0;} 496827bd09bSSatish Balay else if (vals[5]>num_nodes) 497827bd09bSSatish Balay {vals[5]=num_nodes;} 498827bd09bSSatish Balay gs->level = vals[5]; 499827bd09bSSatish Balay 500827bd09bSSatish Balay return(gs); 501827bd09bSSatish Balay } 502827bd09bSSatish Balay 503827bd09bSSatish Balay 504827bd09bSSatish Balay /****************************************************************************** 505827bd09bSSatish Balay Function: gsi_via_bit_mask() 506827bd09bSSatish Balay 507827bd09bSSatish Balay Input : 508827bd09bSSatish Balay Output: 509827bd09bSSatish Balay Return: 510827bd09bSSatish Balay Description: 511827bd09bSSatish Balay 512827bd09bSSatish Balay 513827bd09bSSatish Balay ******************************************************************************/ 5140924e98cSBarry Smith static PetscErrorCode gsi_via_bit_mask(gs_id *gs) 515827bd09bSSatish Balay { 51652f87cdaSBarry Smith PetscInt i, nel, *elms; 51752f87cdaSBarry Smith PetscInt t1; 51852f87cdaSBarry Smith PetscInt **reduce; 51952f87cdaSBarry Smith PetscInt *map; 520827bd09bSSatish Balay 521827bd09bSSatish Balay /* totally local removes ... ct_bits == 0 */ 522827bd09bSSatish Balay get_ngh_buf(gs); 523827bd09bSSatish Balay 524827bd09bSSatish Balay if (gs->level) 525827bd09bSSatish Balay {set_pairwise(gs);} 526827bd09bSSatish Balay 527827bd09bSSatish Balay if (gs->max_left_over) 528827bd09bSSatish Balay {set_tree(gs);} 529827bd09bSSatish Balay 530827bd09bSSatish Balay /* intersection local and pairwise/tree? */ 531827bd09bSSatish Balay gs->num_local_total = gs->num_local; 532827bd09bSSatish Balay gs->gop_local_reduce = gs->local_reduce; 533827bd09bSSatish Balay gs->num_gop_local_reduce = gs->num_local_reduce; 534827bd09bSSatish Balay 535827bd09bSSatish Balay map = gs->companion; 536827bd09bSSatish Balay 537827bd09bSSatish Balay /* is there any local compression */ 538d890fc11SSatish Balay if (!gs->num_local) { 539827bd09bSSatish Balay gs->local_strength = NONE; 540827bd09bSSatish Balay gs->num_local_gop = 0; 541d890fc11SSatish Balay } else { 542827bd09bSSatish Balay /* ok find intersection */ 543827bd09bSSatish Balay map = gs->companion; 544827bd09bSSatish Balay reduce = gs->local_reduce; 545827bd09bSSatish Balay for (i=0, t1=0; i<gs->num_local; i++, reduce++) 546827bd09bSSatish Balay { 547827bd09bSSatish Balay if ((ivec_binary_search(**reduce,gs->pw_elm_list,gs->len_pw_list)>=0) 548827bd09bSSatish Balay || 549827bd09bSSatish Balay ivec_binary_search(**reduce,gs->tree_map_in,gs->tree_map_sz)>=0) 550827bd09bSSatish Balay { 551827bd09bSSatish Balay /* printf("C%d :: i=%d, **reduce=%d\n",my_id,i,**reduce); */ 552827bd09bSSatish Balay t1++; 553827bd09bSSatish Balay if (gs->num_local_reduce[i]<=0) 554388eb383SBarry Smith {SETERRQ(PETSC_ERR_PLIB,"nobody in list?");} 555827bd09bSSatish Balay gs->num_local_reduce[i] *= -1; 556827bd09bSSatish Balay } 557827bd09bSSatish Balay **reduce=map[**reduce]; 558827bd09bSSatish Balay } 559827bd09bSSatish Balay 560827bd09bSSatish Balay /* intersection is empty */ 561827bd09bSSatish Balay if (!t1) 562827bd09bSSatish Balay { 563827bd09bSSatish Balay gs->local_strength = FULL; 564827bd09bSSatish Balay gs->num_local_gop = 0; 565827bd09bSSatish Balay } 566827bd09bSSatish Balay /* intersection not empty */ 567827bd09bSSatish Balay else 568827bd09bSSatish Balay { 569827bd09bSSatish Balay gs->local_strength = PARTIAL; 570827bd09bSSatish Balay SMI_sort((void*)gs->num_local_reduce, (void*)gs->local_reduce, 571827bd09bSSatish Balay gs->num_local + 1, SORT_INT_PTR); 572827bd09bSSatish Balay 573827bd09bSSatish Balay gs->num_local_gop = t1; 574827bd09bSSatish Balay gs->num_local_total = gs->num_local; 575827bd09bSSatish Balay gs->num_local -= t1; 576827bd09bSSatish Balay gs->gop_local_reduce = gs->local_reduce; 577827bd09bSSatish Balay gs->num_gop_local_reduce = gs->num_local_reduce; 578827bd09bSSatish Balay 579827bd09bSSatish Balay for (i=0; i<t1; i++) 580827bd09bSSatish Balay { 581827bd09bSSatish Balay if (gs->num_gop_local_reduce[i]>=0) 582388eb383SBarry Smith {SETERRQ(PETSC_ERR_PLIB,"they aren't negative?");} 583827bd09bSSatish Balay gs->num_gop_local_reduce[i] *= -1; 584827bd09bSSatish Balay gs->local_reduce++; 585827bd09bSSatish Balay gs->num_local_reduce++; 586827bd09bSSatish Balay } 587827bd09bSSatish Balay gs->local_reduce++; 588827bd09bSSatish Balay gs->num_local_reduce++; 589827bd09bSSatish Balay } 590827bd09bSSatish Balay } 591827bd09bSSatish Balay 592827bd09bSSatish Balay elms = gs->pw_elm_list; 593827bd09bSSatish Balay nel = gs->len_pw_list; 594827bd09bSSatish Balay for (i=0; i<nel; i++) 595827bd09bSSatish Balay {elms[i] = map[elms[i]];} 596827bd09bSSatish Balay 597827bd09bSSatish Balay elms = gs->tree_map_in; 598827bd09bSSatish Balay nel = gs->tree_map_sz; 599827bd09bSSatish Balay for (i=0; i<nel; i++) 600827bd09bSSatish Balay {elms[i] = map[elms[i]];} 601827bd09bSSatish Balay 602827bd09bSSatish Balay /* clean up */ 603a501084fSBarry Smith free((void*) gs->local_elms); 604a501084fSBarry Smith free((void*) gs->companion); 605a501084fSBarry Smith free((void*) gs->elms); 606a501084fSBarry Smith free((void*) gs->ngh_buf); 607827bd09bSSatish Balay gs->local_elms = gs->companion = gs->elms = gs->ngh_buf = NULL; 6083fdc5746SBarry Smith PetscFunctionReturn(0); 609827bd09bSSatish Balay } 610827bd09bSSatish Balay 611827bd09bSSatish Balay 612827bd09bSSatish Balay 613827bd09bSSatish Balay /****************************************************************************** 614827bd09bSSatish Balay Function: place_in_tree() 615827bd09bSSatish Balay 616827bd09bSSatish Balay Input : 617827bd09bSSatish Balay Output: 618827bd09bSSatish Balay Return: 619827bd09bSSatish Balay Description: 620827bd09bSSatish Balay 621827bd09bSSatish Balay 622827bd09bSSatish Balay ******************************************************************************/ 62352f87cdaSBarry Smith static PetscErrorCode place_in_tree( PetscInt elm) 624827bd09bSSatish Balay { 62552f87cdaSBarry Smith PetscInt *tp, n; 626827bd09bSSatish Balay 6273fdc5746SBarry Smith PetscFunctionBegin; 628827bd09bSSatish Balay if (ntree==tree_buf_sz) 629827bd09bSSatish Balay { 630827bd09bSSatish Balay if (tree_buf_sz) 631827bd09bSSatish Balay { 632827bd09bSSatish Balay tp = tree_buf; 633827bd09bSSatish Balay n = tree_buf_sz; 634827bd09bSSatish Balay tree_buf_sz<<=1; 63552f87cdaSBarry Smith tree_buf = (PetscInt*)malloc(tree_buf_sz*sizeof(PetscInt)); 636827bd09bSSatish Balay ivec_copy(tree_buf,tp,n); 637a501084fSBarry Smith free(tp); 638827bd09bSSatish Balay } 639827bd09bSSatish Balay else 640827bd09bSSatish Balay { 641827bd09bSSatish Balay tree_buf_sz = TREE_BUF_SZ; 64252f87cdaSBarry Smith tree_buf = (PetscInt*)malloc(tree_buf_sz*sizeof(PetscInt)); 643827bd09bSSatish Balay } 644827bd09bSSatish Balay } 645827bd09bSSatish Balay 646827bd09bSSatish Balay tree_buf[ntree++] = elm; 6473fdc5746SBarry Smith PetscFunctionReturn(0); 648827bd09bSSatish Balay } 649827bd09bSSatish Balay 650827bd09bSSatish Balay 651827bd09bSSatish Balay 652827bd09bSSatish Balay /****************************************************************************** 653827bd09bSSatish Balay Function: get_ngh_buf() 654827bd09bSSatish Balay 655827bd09bSSatish Balay Input : 656827bd09bSSatish Balay Output: 657827bd09bSSatish Balay Return: 658827bd09bSSatish Balay Description: 659827bd09bSSatish Balay 660827bd09bSSatish Balay 661827bd09bSSatish Balay ******************************************************************************/ 6620924e98cSBarry Smith static PetscErrorCode get_ngh_buf(gs_id *gs) 663827bd09bSSatish Balay { 66452f87cdaSBarry Smith PetscInt i, j, npw=0, ntree_map=0; 66552f87cdaSBarry Smith PetscInt p_mask_size, ngh_buf_size, buf_size; 66652f87cdaSBarry Smith PetscInt *p_mask, *sh_proc_mask, *pw_sh_proc_mask; 66752f87cdaSBarry Smith PetscInt *ngh_buf, *buf1, *buf2; 66852f87cdaSBarry Smith PetscInt offset, per_load, num_loads, or_ct, start, end; 66952f87cdaSBarry Smith PetscInt *ptr1, *ptr2, i_start, negl, nel, *elms; 67052f87cdaSBarry Smith PetscInt oper=GL_B_OR; 67152f87cdaSBarry Smith PetscInt *ptr3, *t_mask, level, ct1, ct2; 672827bd09bSSatish Balay 6733fdc5746SBarry Smith PetscFunctionBegin; 674827bd09bSSatish Balay /* to make life easier */ 675827bd09bSSatish Balay nel = gs->nel; 676827bd09bSSatish Balay elms = gs->elms; 677827bd09bSSatish Balay level = gs->level; 678827bd09bSSatish Balay 679827bd09bSSatish Balay /* det #bytes needed for processor bit masks and init w/mask cor. to my_id */ 68052f87cdaSBarry Smith p_mask = (PetscInt*) malloc(p_mask_size=len_bit_mask(num_nodes)); 681827bd09bSSatish Balay set_bit_mask(p_mask,p_mask_size,my_id); 682827bd09bSSatish Balay 683827bd09bSSatish Balay /* allocate space for masks and info bufs */ 68452f87cdaSBarry Smith gs->nghs = sh_proc_mask = (PetscInt*) malloc(p_mask_size); 68552f87cdaSBarry Smith gs->pw_nghs = pw_sh_proc_mask = (PetscInt*) malloc(p_mask_size); 686827bd09bSSatish Balay gs->ngh_buf_sz = ngh_buf_size = p_mask_size*nel; 68752f87cdaSBarry Smith t_mask = (PetscInt*) malloc(p_mask_size); 68852f87cdaSBarry Smith gs->ngh_buf = ngh_buf = (PetscInt*) malloc(ngh_buf_size); 689827bd09bSSatish Balay 690827bd09bSSatish Balay /* comm buffer size ... memory usage bounded by ~2*msg_buf */ 691827bd09bSSatish Balay /* had thought I could exploit rendezvous threshold */ 692827bd09bSSatish Balay 693827bd09bSSatish Balay /* default is one pass */ 694827bd09bSSatish Balay per_load = negl = gs->negl; 695827bd09bSSatish Balay gs->num_loads = num_loads = 1; 696827bd09bSSatish Balay i=p_mask_size*negl; 697827bd09bSSatish Balay 698827bd09bSSatish Balay /* possible overflow on buffer size */ 699827bd09bSSatish Balay /* overflow hack */ 700827bd09bSSatish Balay if (i<0) {i=INT_MAX;} 701827bd09bSSatish Balay 70239945688SSatish Balay buf_size = PetscMin(msg_buf,i); 703827bd09bSSatish Balay 704827bd09bSSatish Balay /* can we do it? */ 705827bd09bSSatish Balay if (p_mask_size>buf_size) 706388eb383SBarry Smith {SETERRQ2(PETSC_ERR_PLIB,"get_ngh_buf() :: buf<pms :: %d>%d\n",p_mask_size,buf_size);} 707827bd09bSSatish Balay 708827bd09bSSatish Balay /* get giop buf space ... make *only* one malloc */ 70952f87cdaSBarry Smith buf1 = (PetscInt*) malloc(buf_size<<1); 710827bd09bSSatish Balay 711827bd09bSSatish Balay /* more than one gior exchange needed? */ 712827bd09bSSatish Balay if (buf_size!=i) 713827bd09bSSatish Balay { 714827bd09bSSatish Balay per_load = buf_size/p_mask_size; 715827bd09bSSatish Balay buf_size = per_load*p_mask_size; 716827bd09bSSatish Balay gs->num_loads = num_loads = negl/per_load + (negl%per_load>0); 717827bd09bSSatish Balay } 718827bd09bSSatish Balay 719827bd09bSSatish Balay 720827bd09bSSatish Balay /* convert buf sizes from #bytes to #ints - 32 bit only! */ 721a501084fSBarry Smith p_mask_size/=sizeof(PetscInt); ngh_buf_size/=sizeof(PetscInt); buf_size/=sizeof(PetscInt); 722827bd09bSSatish Balay 723827bd09bSSatish Balay /* find giop work space */ 724827bd09bSSatish Balay buf2 = buf1+buf_size; 725827bd09bSSatish Balay 726827bd09bSSatish Balay /* hold #ints needed for processor masks */ 727827bd09bSSatish Balay gs->mask_sz=p_mask_size; 728827bd09bSSatish Balay 729827bd09bSSatish Balay /* init buffers */ 730827bd09bSSatish Balay ivec_zero(sh_proc_mask,p_mask_size); 731827bd09bSSatish Balay ivec_zero(pw_sh_proc_mask,p_mask_size); 732827bd09bSSatish Balay ivec_zero(ngh_buf,ngh_buf_size); 733827bd09bSSatish Balay 734827bd09bSSatish Balay /* HACK reset tree info */ 735827bd09bSSatish Balay tree_buf=NULL; 736827bd09bSSatish Balay tree_buf_sz=ntree=0; 737827bd09bSSatish Balay 738827bd09bSSatish Balay /* ok do it */ 739827bd09bSSatish Balay for (ptr1=ngh_buf,ptr2=elms,end=gs->gl_min,or_ct=i=0; or_ct<num_loads; or_ct++) 740827bd09bSSatish Balay { 741827bd09bSSatish Balay /* identity for bitwise or is 000...000 */ 742827bd09bSSatish Balay ivec_zero(buf1,buf_size); 743827bd09bSSatish Balay 744827bd09bSSatish Balay /* load msg buffer */ 745827bd09bSSatish Balay for (start=end,end+=per_load,i_start=i; (offset=*ptr2)<end; i++, ptr2++) 746827bd09bSSatish Balay { 747827bd09bSSatish Balay offset = (offset-start)*p_mask_size; 748827bd09bSSatish Balay ivec_copy(buf1+offset,p_mask,p_mask_size); 749827bd09bSSatish Balay } 750827bd09bSSatish Balay 751827bd09bSSatish Balay /* GLOBAL: pass buffer */ 752827bd09bSSatish Balay giop(buf1,buf2,buf_size,&oper); 753827bd09bSSatish Balay 754827bd09bSSatish Balay 755827bd09bSSatish Balay /* unload buffer into ngh_buf */ 756827bd09bSSatish Balay ptr2=(elms+i_start); 757827bd09bSSatish Balay for(ptr3=buf1,j=start; j<end; ptr3+=p_mask_size,j++) 758827bd09bSSatish Balay { 759827bd09bSSatish Balay /* I own it ... may have to pairwise it */ 760827bd09bSSatish Balay if (j==*ptr2) 761827bd09bSSatish Balay { 762827bd09bSSatish Balay /* do i share it w/anyone? */ 763a501084fSBarry Smith ct1 = ct_bits((char *)ptr3,p_mask_size*sizeof(PetscInt)); 764827bd09bSSatish Balay /* guess not */ 765827bd09bSSatish Balay if (ct1<2) 766827bd09bSSatish Balay {ptr2++; ptr1+=p_mask_size; continue;} 767827bd09bSSatish Balay 768827bd09bSSatish Balay /* i do ... so keep info and turn off my bit */ 769827bd09bSSatish Balay ivec_copy(ptr1,ptr3,p_mask_size); 770827bd09bSSatish Balay ivec_xor(ptr1,p_mask,p_mask_size); 771827bd09bSSatish Balay ivec_or(sh_proc_mask,ptr1,p_mask_size); 772827bd09bSSatish Balay 773827bd09bSSatish Balay /* is it to be done pairwise? */ 774827bd09bSSatish Balay if (--ct1<=level) 775827bd09bSSatish Balay { 776827bd09bSSatish Balay npw++; 777827bd09bSSatish Balay 778827bd09bSSatish Balay /* turn on high bit to indicate pw need to process */ 779827bd09bSSatish Balay *ptr2++ |= TOP_BIT; 780827bd09bSSatish Balay ivec_or(pw_sh_proc_mask,ptr1,p_mask_size); 781827bd09bSSatish Balay ptr1+=p_mask_size; 782827bd09bSSatish Balay continue; 783827bd09bSSatish Balay } 784827bd09bSSatish Balay 785827bd09bSSatish Balay /* get set for next and note that I have a tree contribution */ 786827bd09bSSatish Balay /* could save exact elm index for tree here -> save a search */ 787827bd09bSSatish Balay ptr2++; ptr1+=p_mask_size; ntree_map++; 788827bd09bSSatish Balay } 789827bd09bSSatish Balay /* i don't but still might be involved in tree */ 790827bd09bSSatish Balay else 791827bd09bSSatish Balay { 792827bd09bSSatish Balay 793827bd09bSSatish Balay /* shared by how many? */ 794a501084fSBarry Smith ct1 = ct_bits((char *)ptr3,p_mask_size*sizeof(PetscInt)); 795827bd09bSSatish Balay 796827bd09bSSatish Balay /* none! */ 797827bd09bSSatish Balay if (ct1<2) 798827bd09bSSatish Balay {continue;} 799827bd09bSSatish Balay 800827bd09bSSatish Balay /* is it going to be done pairwise? but not by me of course!*/ 801827bd09bSSatish Balay if (--ct1<=level) 802827bd09bSSatish Balay {continue;} 803827bd09bSSatish Balay } 804827bd09bSSatish Balay /* LATER we're going to have to process it NOW */ 805827bd09bSSatish Balay /* nope ... tree it */ 806827bd09bSSatish Balay place_in_tree(j); 807827bd09bSSatish Balay } 808827bd09bSSatish Balay } 809827bd09bSSatish Balay 810a501084fSBarry Smith free((void*)t_mask); 811a501084fSBarry Smith free((void*)buf1); 812827bd09bSSatish Balay 813827bd09bSSatish Balay gs->len_pw_list=npw; 814a501084fSBarry Smith gs->num_nghs = ct_bits((char *)sh_proc_mask,p_mask_size*sizeof(PetscInt)); 815827bd09bSSatish Balay 816827bd09bSSatish Balay /* expand from bit mask list to int list and save ngh list */ 81752f87cdaSBarry Smith gs->nghs = (PetscInt*) malloc(gs->num_nghs * sizeof(PetscInt)); 818a501084fSBarry Smith bm_to_proc((char *)sh_proc_mask,p_mask_size*sizeof(PetscInt),gs->nghs); 819827bd09bSSatish Balay 820a501084fSBarry Smith gs->num_pw_nghs = ct_bits((char *)pw_sh_proc_mask,p_mask_size*sizeof(PetscInt)); 821827bd09bSSatish Balay 822827bd09bSSatish Balay oper = GL_MAX; 823827bd09bSSatish Balay ct1 = gs->num_nghs; 824827bd09bSSatish Balay giop(&ct1,&ct2,1,&oper); 825827bd09bSSatish Balay gs->max_nghs = ct1; 826827bd09bSSatish Balay 827827bd09bSSatish Balay gs->tree_map_sz = ntree_map; 828827bd09bSSatish Balay gs->max_left_over=ntree; 829827bd09bSSatish Balay 830a501084fSBarry Smith free((void*)p_mask); 831a501084fSBarry Smith free((void*)sh_proc_mask); 8323fdc5746SBarry Smith PetscFunctionReturn(0); 833827bd09bSSatish Balay } 834827bd09bSSatish Balay 835827bd09bSSatish Balay 836827bd09bSSatish Balay 837827bd09bSSatish Balay 838827bd09bSSatish Balay 839827bd09bSSatish Balay /****************************************************************************** 840827bd09bSSatish Balay Function: pairwise_init() 841827bd09bSSatish Balay 842827bd09bSSatish Balay Input : 843827bd09bSSatish Balay Output: 844827bd09bSSatish Balay Return: 845827bd09bSSatish Balay Description: 846827bd09bSSatish Balay 847827bd09bSSatish Balay if an element is shared by fewer that level# of nodes do pairwise exch 848827bd09bSSatish Balay ******************************************************************************/ 8490924e98cSBarry Smith static PetscErrorCode set_pairwise(gs_id *gs) 850827bd09bSSatish Balay { 85152f87cdaSBarry Smith PetscInt i, j; 85252f87cdaSBarry Smith PetscInt p_mask_size; 85352f87cdaSBarry Smith PetscInt *p_mask, *sh_proc_mask, *tmp_proc_mask; 85452f87cdaSBarry Smith PetscInt *ngh_buf, *buf2; 85552f87cdaSBarry Smith PetscInt offset; 85652f87cdaSBarry Smith PetscInt *msg_list, *msg_size, **msg_nodes, nprs; 85752f87cdaSBarry Smith PetscInt *pairwise_elm_list, len_pair_list=0; 85852f87cdaSBarry Smith PetscInt *iptr, t1, i_start, nel, *elms; 85952f87cdaSBarry Smith PetscInt ct; 860827bd09bSSatish Balay 861827bd09bSSatish Balay 8623fdc5746SBarry Smith PetscFunctionBegin; 863827bd09bSSatish Balay /* to make life easier */ 864827bd09bSSatish Balay nel = gs->nel; 865827bd09bSSatish Balay elms = gs->elms; 866827bd09bSSatish Balay ngh_buf = gs->ngh_buf; 867827bd09bSSatish Balay sh_proc_mask = gs->pw_nghs; 868827bd09bSSatish Balay 869827bd09bSSatish Balay /* need a few temp masks */ 870827bd09bSSatish Balay p_mask_size = len_bit_mask(num_nodes); 87152f87cdaSBarry Smith p_mask = (PetscInt*) malloc(p_mask_size); 87252f87cdaSBarry Smith tmp_proc_mask = (PetscInt*) malloc(p_mask_size); 873827bd09bSSatish Balay 874827bd09bSSatish Balay /* set mask to my my_id's bit mask */ 875827bd09bSSatish Balay set_bit_mask(p_mask,p_mask_size,my_id); 876827bd09bSSatish Balay 877a501084fSBarry Smith p_mask_size /= sizeof(PetscInt); 878827bd09bSSatish Balay 879827bd09bSSatish Balay len_pair_list=gs->len_pw_list; 88052f87cdaSBarry Smith gs->pw_elm_list=pairwise_elm_list=(PetscInt*)malloc((len_pair_list+1)*sizeof(PetscInt)); 881827bd09bSSatish Balay 882827bd09bSSatish Balay /* how many processors (nghs) do we have to exchange with? */ 883a501084fSBarry Smith nprs=gs->num_pairs=ct_bits((char *)sh_proc_mask,p_mask_size*sizeof(PetscInt)); 884827bd09bSSatish Balay 885827bd09bSSatish Balay 886827bd09bSSatish Balay /* allocate space for gs_gop() info */ 88752f87cdaSBarry Smith gs->pair_list = msg_list = (PetscInt *) malloc(sizeof(PetscInt)*nprs); 88852f87cdaSBarry Smith gs->msg_sizes = msg_size = (PetscInt *) malloc(sizeof(PetscInt)*nprs); 88952f87cdaSBarry Smith gs->node_list = msg_nodes = (PetscInt **) malloc(sizeof(PetscInt*)*(nprs+1)); 890827bd09bSSatish Balay 891827bd09bSSatish Balay /* init msg_size list */ 892827bd09bSSatish Balay ivec_zero(msg_size,nprs); 893827bd09bSSatish Balay 894827bd09bSSatish Balay /* expand from bit mask list to int list */ 895a501084fSBarry Smith bm_to_proc((char *)sh_proc_mask,p_mask_size*sizeof(PetscInt),msg_list); 896827bd09bSSatish Balay 897827bd09bSSatish Balay /* keep list of elements being handled pairwise */ 898827bd09bSSatish Balay for (i=j=0;i<nel;i++) 899827bd09bSSatish Balay { 900827bd09bSSatish Balay if (elms[i] & TOP_BIT) 901827bd09bSSatish Balay {elms[i] ^= TOP_BIT; pairwise_elm_list[j++] = i;} 902827bd09bSSatish Balay } 903827bd09bSSatish Balay pairwise_elm_list[j] = -1; 904827bd09bSSatish Balay 905a501084fSBarry Smith gs->msg_ids_out = (MPI_Request *) malloc(sizeof(MPI_Request)*(nprs+1)); 906827bd09bSSatish Balay gs->msg_ids_out[nprs] = MPI_REQUEST_NULL; 907a501084fSBarry Smith gs->msg_ids_in = (MPI_Request *) malloc(sizeof(MPI_Request)*(nprs+1)); 908827bd09bSSatish Balay gs->msg_ids_in[nprs] = MPI_REQUEST_NULL; 909a501084fSBarry Smith gs->pw_vals = (PetscScalar *) malloc(sizeof(PetscScalar)*len_pair_list*vec_sz); 910827bd09bSSatish Balay 911827bd09bSSatish Balay /* find who goes to each processor */ 912827bd09bSSatish Balay for (i_start=i=0;i<nprs;i++) 913827bd09bSSatish Balay { 914827bd09bSSatish Balay /* processor i's mask */ 915a501084fSBarry Smith set_bit_mask(p_mask,p_mask_size*sizeof(PetscInt),msg_list[i]); 916827bd09bSSatish Balay 917827bd09bSSatish Balay /* det # going to processor i */ 918827bd09bSSatish Balay for (ct=j=0;j<len_pair_list;j++) 919827bd09bSSatish Balay { 920827bd09bSSatish Balay buf2 = ngh_buf+(pairwise_elm_list[j]*p_mask_size); 921827bd09bSSatish Balay ivec_and3(tmp_proc_mask,p_mask,buf2,p_mask_size); 922a501084fSBarry Smith if (ct_bits((char *)tmp_proc_mask,p_mask_size*sizeof(PetscInt))) 923827bd09bSSatish Balay {ct++;} 924827bd09bSSatish Balay } 925827bd09bSSatish Balay msg_size[i] = ct; 92639945688SSatish Balay i_start = PetscMax(i_start,ct); 927827bd09bSSatish Balay 928827bd09bSSatish Balay /*space to hold nodes in message to first neighbor */ 92952f87cdaSBarry Smith msg_nodes[i] = iptr = (PetscInt*) malloc(sizeof(PetscInt)*(ct+1)); 930827bd09bSSatish Balay 931827bd09bSSatish Balay for (j=0;j<len_pair_list;j++) 932827bd09bSSatish Balay { 933827bd09bSSatish Balay buf2 = ngh_buf+(pairwise_elm_list[j]*p_mask_size); 934827bd09bSSatish Balay ivec_and3(tmp_proc_mask,p_mask,buf2,p_mask_size); 935a501084fSBarry Smith if (ct_bits((char *)tmp_proc_mask,p_mask_size*sizeof(PetscInt))) 936827bd09bSSatish Balay {*iptr++ = j;} 937827bd09bSSatish Balay } 938827bd09bSSatish Balay *iptr = -1; 939827bd09bSSatish Balay } 940827bd09bSSatish Balay msg_nodes[nprs] = NULL; 941827bd09bSSatish Balay 942827bd09bSSatish Balay j=gs->loc_node_pairs=i_start; 943827bd09bSSatish Balay t1 = GL_MAX; 944827bd09bSSatish Balay giop(&i_start,&offset,1,&t1); 945827bd09bSSatish Balay gs->max_node_pairs = i_start; 946827bd09bSSatish Balay 947827bd09bSSatish Balay i_start=j; 948827bd09bSSatish Balay t1 = GL_MIN; 949827bd09bSSatish Balay giop(&i_start,&offset,1,&t1); 950827bd09bSSatish Balay gs->min_node_pairs = i_start; 951827bd09bSSatish Balay 952827bd09bSSatish Balay i_start=j; 953827bd09bSSatish Balay t1 = GL_ADD; 954827bd09bSSatish Balay giop(&i_start,&offset,1,&t1); 955827bd09bSSatish Balay gs->avg_node_pairs = i_start/num_nodes + 1; 956827bd09bSSatish Balay 957827bd09bSSatish Balay i_start=nprs; 958827bd09bSSatish Balay t1 = GL_MAX; 959827bd09bSSatish Balay giop(&i_start,&offset,1,&t1); 960827bd09bSSatish Balay gs->max_pairs = i_start; 961827bd09bSSatish Balay 962827bd09bSSatish Balay 963827bd09bSSatish Balay /* remap pairwise in tail of gsi_via_bit_mask() */ 964827bd09bSSatish Balay gs->msg_total = ivec_sum(gs->msg_sizes,nprs); 965a501084fSBarry Smith gs->out = (PetscScalar *) malloc(sizeof(PetscScalar)*gs->msg_total*vec_sz); 966a501084fSBarry Smith gs->in = (PetscScalar *) malloc(sizeof(PetscScalar)*gs->msg_total*vec_sz); 967827bd09bSSatish Balay 968827bd09bSSatish Balay /* reset malloc pool */ 969a501084fSBarry Smith free((void*)p_mask); 970a501084fSBarry Smith free((void*)tmp_proc_mask); 9713fdc5746SBarry Smith PetscFunctionReturn(0); 972827bd09bSSatish Balay } 973827bd09bSSatish Balay 974827bd09bSSatish Balay 975827bd09bSSatish Balay 976827bd09bSSatish Balay /****************************************************************************** 977827bd09bSSatish Balay Function: set_tree() 978827bd09bSSatish Balay 979827bd09bSSatish Balay Input : 980827bd09bSSatish Balay Output: 981827bd09bSSatish Balay Return: 982827bd09bSSatish Balay Description: 983827bd09bSSatish Balay 984827bd09bSSatish Balay to do pruned tree just save ngh buf copy for each one and decode here! 985827bd09bSSatish Balay ******************************************************************************/ 9860924e98cSBarry Smith static PetscErrorCode set_tree(gs_id *gs) 987827bd09bSSatish Balay { 98852f87cdaSBarry Smith PetscInt i, j, n, nel; 98952f87cdaSBarry Smith PetscInt *iptr_in, *iptr_out, *tree_elms, *elms; 990827bd09bSSatish Balay 9913fdc5746SBarry Smith PetscFunctionBegin; 992827bd09bSSatish Balay /* local work ptrs */ 993827bd09bSSatish Balay elms = gs->elms; 994827bd09bSSatish Balay nel = gs->nel; 995827bd09bSSatish Balay 996827bd09bSSatish Balay /* how many via tree */ 997827bd09bSSatish Balay gs->tree_nel = n = ntree; 998827bd09bSSatish Balay gs->tree_elms = tree_elms = iptr_in = tree_buf; 999a501084fSBarry Smith gs->tree_buf = (PetscScalar *) malloc(sizeof(PetscScalar)*n*vec_sz); 1000a501084fSBarry Smith gs->tree_work = (PetscScalar *) malloc(sizeof(PetscScalar)*n*vec_sz); 1001827bd09bSSatish Balay j=gs->tree_map_sz; 100252f87cdaSBarry Smith gs->tree_map_in = iptr_in = (PetscInt*) malloc(sizeof(PetscInt)*(j+1)); 100352f87cdaSBarry Smith gs->tree_map_out = iptr_out = (PetscInt*) malloc(sizeof(PetscInt)*(j+1)); 1004827bd09bSSatish Balay 1005827bd09bSSatish Balay /* search the longer of the two lists */ 1006827bd09bSSatish Balay /* note ... could save this info in get_ngh_buf and save searches */ 1007827bd09bSSatish Balay if (n<=nel) 1008827bd09bSSatish Balay { 1009827bd09bSSatish Balay /* bijective fct w/remap - search elm list */ 1010827bd09bSSatish Balay for (i=0; i<n; i++) 1011827bd09bSSatish Balay { 1012827bd09bSSatish Balay if ((j=ivec_binary_search(*tree_elms++,elms,nel))>=0) 1013827bd09bSSatish Balay {*iptr_in++ = j; *iptr_out++ = i;} 1014827bd09bSSatish Balay } 1015827bd09bSSatish Balay } 1016827bd09bSSatish Balay else 1017827bd09bSSatish Balay { 1018827bd09bSSatish Balay for (i=0; i<nel; i++) 1019827bd09bSSatish Balay { 1020827bd09bSSatish Balay if ((j=ivec_binary_search(*elms++,tree_elms,n))>=0) 1021827bd09bSSatish Balay {*iptr_in++ = i; *iptr_out++ = j;} 1022827bd09bSSatish Balay } 1023827bd09bSSatish Balay } 1024827bd09bSSatish Balay 1025827bd09bSSatish Balay /* sentinel */ 1026827bd09bSSatish Balay *iptr_in = *iptr_out = -1; 10273fdc5746SBarry Smith PetscFunctionReturn(0); 1028827bd09bSSatish Balay } 1029827bd09bSSatish Balay 1030827bd09bSSatish Balay 1031827bd09bSSatish Balay /****************************************************************************** 1032827bd09bSSatish Balay Function: gather_scatter 1033827bd09bSSatish Balay 1034827bd09bSSatish Balay Input : 1035827bd09bSSatish Balay Output: 1036827bd09bSSatish Balay Return: 1037827bd09bSSatish Balay Description: 1038827bd09bSSatish Balay ******************************************************************************/ 10390924e98cSBarry Smith static PetscErrorCode gs_gop_local_out( gs_id *gs, PetscScalar *vals) 1040827bd09bSSatish Balay { 104152f87cdaSBarry Smith PetscInt *num, *map, **reduce; 1042a501084fSBarry Smith PetscScalar tmp; 1043827bd09bSSatish Balay 10443fdc5746SBarry Smith PetscFunctionBegin; 1045827bd09bSSatish Balay num = gs->num_gop_local_reduce; 1046827bd09bSSatish Balay reduce = gs->gop_local_reduce; 1047827bd09bSSatish Balay while ((map = *reduce++)) 1048827bd09bSSatish Balay { 1049827bd09bSSatish Balay /* wall */ 1050827bd09bSSatish Balay if (*num == 2) 1051827bd09bSSatish Balay { 1052827bd09bSSatish Balay num ++; 1053827bd09bSSatish Balay vals[map[1]] = vals[map[0]]; 1054827bd09bSSatish Balay } 1055827bd09bSSatish Balay /* corner shared by three elements */ 1056827bd09bSSatish Balay else if (*num == 3) 1057827bd09bSSatish Balay { 1058827bd09bSSatish Balay num ++; 1059827bd09bSSatish Balay vals[map[2]] = vals[map[1]] = vals[map[0]]; 1060827bd09bSSatish Balay } 1061827bd09bSSatish Balay /* corner shared by four elements */ 1062827bd09bSSatish Balay else if (*num == 4) 1063827bd09bSSatish Balay { 1064827bd09bSSatish Balay num ++; 1065827bd09bSSatish Balay vals[map[3]] = vals[map[2]] = vals[map[1]] = vals[map[0]]; 1066827bd09bSSatish Balay } 1067827bd09bSSatish Balay /* general case ... odd geoms ... 3D*/ 1068827bd09bSSatish Balay else 1069827bd09bSSatish Balay { 1070827bd09bSSatish Balay num++; 1071827bd09bSSatish Balay tmp = *(vals + *map++); 1072827bd09bSSatish Balay while (*map >= 0) 1073827bd09bSSatish Balay {*(vals + *map++) = tmp;} 1074827bd09bSSatish Balay } 1075827bd09bSSatish Balay } 10763fdc5746SBarry Smith PetscFunctionReturn(0); 1077827bd09bSSatish Balay } 1078827bd09bSSatish Balay 1079827bd09bSSatish Balay 1080827bd09bSSatish Balay 1081827bd09bSSatish Balay /****************************************************************************** 1082827bd09bSSatish Balay Function: gather_scatter 1083827bd09bSSatish Balay 1084827bd09bSSatish Balay Input : 1085827bd09bSSatish Balay Output: 1086827bd09bSSatish Balay Return: 1087827bd09bSSatish Balay Description: 1088827bd09bSSatish Balay ******************************************************************************/ 10890924e98cSBarry Smith PetscErrorCode gs_gop_binary(gs_ADT gs, PetscScalar *vals, rbfp fct) 1090827bd09bSSatish Balay { 10913fdc5746SBarry Smith PetscFunctionBegin; 1092827bd09bSSatish Balay /* local only operations!!! */ 1093827bd09bSSatish Balay if (gs->num_local) 1094827bd09bSSatish Balay {gs_gop_local_binary(gs,vals,fct);} 1095827bd09bSSatish Balay 1096827bd09bSSatish Balay /* if intersection tree/pairwise and local isn't empty */ 1097827bd09bSSatish Balay if (gs->num_local_gop) 1098827bd09bSSatish Balay { 1099827bd09bSSatish Balay gs_gop_local_in_binary(gs,vals,fct); 1100827bd09bSSatish Balay 1101827bd09bSSatish Balay /* pairwise */ 1102827bd09bSSatish Balay if (gs->num_pairs) 1103827bd09bSSatish Balay {gs_gop_pairwise_binary(gs,vals,fct);} 1104827bd09bSSatish Balay 1105827bd09bSSatish Balay /* tree */ 1106827bd09bSSatish Balay else if (gs->max_left_over) 1107827bd09bSSatish Balay {gs_gop_tree_binary(gs,vals,fct);} 1108827bd09bSSatish Balay 1109827bd09bSSatish Balay gs_gop_local_out(gs,vals); 1110827bd09bSSatish Balay } 1111827bd09bSSatish Balay /* if intersection tree/pairwise and local is empty */ 1112827bd09bSSatish Balay else 1113827bd09bSSatish Balay { 1114827bd09bSSatish Balay /* pairwise */ 1115827bd09bSSatish Balay if (gs->num_pairs) 1116827bd09bSSatish Balay {gs_gop_pairwise_binary(gs,vals,fct);} 1117827bd09bSSatish Balay 1118827bd09bSSatish Balay /* tree */ 1119827bd09bSSatish Balay else if (gs->max_left_over) 1120827bd09bSSatish Balay {gs_gop_tree_binary(gs,vals,fct);} 1121827bd09bSSatish Balay } 11223fdc5746SBarry Smith PetscFunctionReturn(0); 1123827bd09bSSatish Balay } 1124827bd09bSSatish Balay 1125827bd09bSSatish Balay 1126827bd09bSSatish Balay 1127827bd09bSSatish Balay /****************************************************************************** 1128827bd09bSSatish Balay Function: gather_scatter 1129827bd09bSSatish Balay 1130827bd09bSSatish Balay Input : 1131827bd09bSSatish Balay Output: 1132827bd09bSSatish Balay Return: 1133827bd09bSSatish Balay Description: 1134827bd09bSSatish Balay ******************************************************************************/ 11350924e98cSBarry Smith static PetscErrorCode gs_gop_local_binary( gs_id *gs, PetscScalar *vals, rbfp fct) 1136827bd09bSSatish Balay { 113752f87cdaSBarry Smith PetscInt *num, *map, **reduce; 1138a501084fSBarry Smith PetscScalar tmp; 1139827bd09bSSatish Balay 11403fdc5746SBarry Smith PetscFunctionBegin; 1141827bd09bSSatish Balay num = gs->num_local_reduce; 1142827bd09bSSatish Balay reduce = gs->local_reduce; 1143827bd09bSSatish Balay while ((map = *reduce)) 1144827bd09bSSatish Balay { 1145827bd09bSSatish Balay num ++; 1146827bd09bSSatish Balay (*fct)(&tmp,NULL,1); 1147827bd09bSSatish Balay /* tmp = 0.0; */ 1148827bd09bSSatish Balay while (*map >= 0) 1149827bd09bSSatish Balay {(*fct)(&tmp,(vals + *map),1); map++;} 1150827bd09bSSatish Balay /* {tmp = (*fct)(tmp,*(vals + *map)); map++;} */ 1151827bd09bSSatish Balay 1152827bd09bSSatish Balay map = *reduce++; 1153827bd09bSSatish Balay while (*map >= 0) 1154827bd09bSSatish Balay {*(vals + *map++) = tmp;} 1155827bd09bSSatish Balay } 11563fdc5746SBarry Smith PetscFunctionReturn(0); 1157827bd09bSSatish Balay } 1158827bd09bSSatish Balay 1159827bd09bSSatish Balay 1160827bd09bSSatish Balay 1161827bd09bSSatish Balay /****************************************************************************** 1162827bd09bSSatish Balay Function: gather_scatter 1163827bd09bSSatish Balay 1164827bd09bSSatish Balay Input : 1165827bd09bSSatish Balay Output: 1166827bd09bSSatish Balay Return: 1167827bd09bSSatish Balay Description: 1168827bd09bSSatish Balay ******************************************************************************/ 11690924e98cSBarry Smith static PetscErrorCode gs_gop_local_in_binary( gs_id *gs, PetscScalar *vals, rbfp fct) 1170827bd09bSSatish Balay { 117152f87cdaSBarry Smith PetscInt *num, *map, **reduce; 1172a501084fSBarry Smith PetscScalar *base; 1173827bd09bSSatish Balay 11743fdc5746SBarry Smith PetscFunctionBegin; 1175827bd09bSSatish Balay num = gs->num_gop_local_reduce; 1176827bd09bSSatish Balay 1177827bd09bSSatish Balay reduce = gs->gop_local_reduce; 1178827bd09bSSatish Balay while ((map = *reduce++)) 1179827bd09bSSatish Balay { 1180827bd09bSSatish Balay num++; 1181827bd09bSSatish Balay base = vals + *map++; 1182827bd09bSSatish Balay while (*map >= 0) 1183827bd09bSSatish Balay {(*fct)(base,(vals + *map),1); map++;} 1184827bd09bSSatish Balay } 11853fdc5746SBarry Smith PetscFunctionReturn(0); 1186827bd09bSSatish Balay } 1187827bd09bSSatish Balay 1188827bd09bSSatish Balay 1189827bd09bSSatish Balay 1190827bd09bSSatish Balay /****************************************************************************** 1191827bd09bSSatish Balay Function: gather_scatter 1192827bd09bSSatish Balay 1193827bd09bSSatish Balay VERSION 3 :: 1194827bd09bSSatish Balay 1195827bd09bSSatish Balay Input : 1196827bd09bSSatish Balay Output: 1197827bd09bSSatish Balay Return: 1198827bd09bSSatish Balay Description: 1199827bd09bSSatish Balay ******************************************************************************/ 12007b1ae94cSBarry Smith static PetscErrorCode gs_gop_pairwise_binary( gs_id *gs, PetscScalar *in_vals,rbfp fct) 1201827bd09bSSatish Balay { 1202a501084fSBarry Smith PetscScalar *dptr1, *dptr2, *dptr3, *in1, *in2; 120352f87cdaSBarry Smith PetscInt *iptr, *msg_list, *msg_size, **msg_nodes; 120452f87cdaSBarry Smith PetscInt *pw, *list, *size, **nodes; 1205827bd09bSSatish Balay MPI_Request *msg_ids_in, *msg_ids_out, *ids_in, *ids_out; 1206827bd09bSSatish Balay MPI_Status status; 12073fdc5746SBarry Smith PetscErrorCode ierr; 1208827bd09bSSatish Balay 12093fdc5746SBarry Smith PetscFunctionBegin; 1210a501084fSBarry Smith /* strip and load s */ 1211827bd09bSSatish Balay msg_list =list = gs->pair_list; 1212827bd09bSSatish Balay msg_size =size = gs->msg_sizes; 1213827bd09bSSatish Balay msg_nodes=nodes = gs->node_list; 1214827bd09bSSatish Balay iptr=pw = gs->pw_elm_list; 1215827bd09bSSatish Balay dptr1=dptr3 = gs->pw_vals; 1216827bd09bSSatish Balay msg_ids_in = ids_in = gs->msg_ids_in; 1217827bd09bSSatish Balay msg_ids_out = ids_out = gs->msg_ids_out; 1218827bd09bSSatish Balay dptr2 = gs->out; 1219827bd09bSSatish Balay in1=in2 = gs->in; 1220827bd09bSSatish Balay 1221827bd09bSSatish Balay /* post the receives */ 1222827bd09bSSatish Balay /* msg_nodes=nodes; */ 1223827bd09bSSatish Balay do 1224827bd09bSSatish Balay { 1225827bd09bSSatish Balay /* Should MPI_ANY_SOURCE be replaced by *list ? In that case do the 1226827bd09bSSatish Balay second one *list and do list++ afterwards */ 12273fdc5746SBarry Smith ierr = MPI_Irecv(in1, *size, MPIU_SCALAR, MPI_ANY_SOURCE, MSGTAG1 + *list++, gs->gs_comm, msg_ids_in++);CHKERRQ(ierr); 1228827bd09bSSatish Balay in1 += *size++; 1229827bd09bSSatish Balay } 1230827bd09bSSatish Balay while (*++msg_nodes); 1231827bd09bSSatish Balay msg_nodes=nodes; 1232827bd09bSSatish Balay 1233827bd09bSSatish Balay /* load gs values into in out gs buffers */ 1234827bd09bSSatish Balay while (*iptr >= 0) 1235827bd09bSSatish Balay {*dptr3++ = *(in_vals + *iptr++);} 1236827bd09bSSatish Balay 1237827bd09bSSatish Balay /* load out buffers and post the sends */ 1238827bd09bSSatish Balay while ((iptr = *msg_nodes++)) 1239827bd09bSSatish Balay { 1240827bd09bSSatish Balay dptr3 = dptr2; 1241827bd09bSSatish Balay while (*iptr >= 0) 1242827bd09bSSatish Balay {*dptr2++ = *(dptr1 + *iptr++);} 1243827bd09bSSatish Balay /* CHECK PERSISTENT COMMS MODE FOR ALL THIS STUFF */ 1244827bd09bSSatish Balay /* is msg_ids_out++ correct? */ 12453fdc5746SBarry Smith ierr = MPI_Isend(dptr3, *msg_size++, MPIU_SCALAR, *msg_list++, MSGTAG1+my_id, gs->gs_comm, msg_ids_out++);CHKERRQ(ierr); 1246827bd09bSSatish Balay } 1247827bd09bSSatish Balay 1248827bd09bSSatish Balay if (gs->max_left_over) 1249827bd09bSSatish Balay {gs_gop_tree_binary(gs,in_vals,fct);} 1250827bd09bSSatish Balay 1251827bd09bSSatish Balay /* process the received data */ 1252827bd09bSSatish Balay msg_nodes=nodes; 1253827bd09bSSatish Balay while ((iptr = *nodes++)) 1254827bd09bSSatish Balay { 1255827bd09bSSatish Balay /* Should I check the return value of MPI_Wait() or status? */ 1256827bd09bSSatish Balay /* Can this loop be replaced by a call to MPI_Waitall()? */ 12573fdc5746SBarry Smith ierr = MPI_Wait(ids_in++, &status);CHKERRQ(ierr); 1258827bd09bSSatish Balay while (*iptr >= 0) 1259827bd09bSSatish Balay {(*fct)((dptr1 + *iptr),in2,1); iptr++; in2++;} 1260827bd09bSSatish Balay /* {*(dptr1 + *iptr) = (*fct)(*(dptr1 + *iptr),*in2); iptr++; in2++;} */ 1261827bd09bSSatish Balay } 1262827bd09bSSatish Balay 1263827bd09bSSatish Balay /* replace vals */ 1264827bd09bSSatish Balay while (*pw >= 0) 1265827bd09bSSatish Balay {*(in_vals + *pw++) = *dptr1++;} 1266827bd09bSSatish Balay 1267827bd09bSSatish Balay /* clear isend message handles */ 1268827bd09bSSatish Balay /* This changed for clarity though it could be the same */ 1269827bd09bSSatish Balay while (*msg_nodes++) 1270827bd09bSSatish Balay /* Should I check the return value of MPI_Wait() or status? */ 1271827bd09bSSatish Balay /* Can this loop be replaced by a call to MPI_Waitall()? */ 12723fdc5746SBarry Smith {ierr = MPI_Wait(ids_out++, &status);CHKERRQ(ierr);} 12733fdc5746SBarry Smith PetscFunctionReturn(0); 1274827bd09bSSatish Balay } 1275827bd09bSSatish Balay 1276827bd09bSSatish Balay 1277827bd09bSSatish Balay 1278827bd09bSSatish Balay /****************************************************************************** 1279827bd09bSSatish Balay Function: gather_scatter 1280827bd09bSSatish Balay 1281827bd09bSSatish Balay Input : 1282827bd09bSSatish Balay Output: 1283827bd09bSSatish Balay Return: 1284827bd09bSSatish Balay Description: 1285827bd09bSSatish Balay ******************************************************************************/ 12860924e98cSBarry Smith static PetscErrorCode gs_gop_tree_binary(gs_id *gs, PetscScalar *vals, rbfp fct) 1287827bd09bSSatish Balay { 128852f87cdaSBarry Smith PetscInt size; 128952f87cdaSBarry Smith PetscInt *in, *out; 1290a501084fSBarry Smith PetscScalar *buf, *work; 1291827bd09bSSatish Balay 12923fdc5746SBarry Smith PetscFunctionBegin; 1293827bd09bSSatish Balay in = gs->tree_map_in; 1294827bd09bSSatish Balay out = gs->tree_map_out; 1295827bd09bSSatish Balay buf = gs->tree_buf; 1296827bd09bSSatish Balay work = gs->tree_work; 1297827bd09bSSatish Balay size = gs->tree_nel; 1298827bd09bSSatish Balay 1299827bd09bSSatish Balay /* load vals vector w/identity */ 1300827bd09bSSatish Balay (*fct)(buf,NULL,size); 1301827bd09bSSatish Balay 1302827bd09bSSatish Balay /* load my contribution into val vector */ 13037b1ae94cSBarry Smith while (*in >= 0) { 13047b1ae94cSBarry Smith (*fct)((buf + *out++),(vals + *in++),-1); 13057b1ae94cSBarry Smith } 1306827bd09bSSatish Balay 1307a501084fSBarry Smith gfop(buf,work,size,(vbfp)fct,MPIU_SCALAR,0); 1308827bd09bSSatish Balay 1309827bd09bSSatish Balay in = gs->tree_map_in; 1310827bd09bSSatish Balay out = gs->tree_map_out; 13117b1ae94cSBarry Smith while (*in >= 0) { 13127b1ae94cSBarry Smith (*fct)((vals + *in++),(buf + *out++),-1); 13137b1ae94cSBarry Smith } 13143fdc5746SBarry Smith PetscFunctionReturn(0); 1315827bd09bSSatish Balay } 1316827bd09bSSatish Balay 1317827bd09bSSatish Balay 1318827bd09bSSatish Balay 1319827bd09bSSatish Balay 1320827bd09bSSatish Balay /****************************************************************************** 1321827bd09bSSatish Balay Function: gather_scatter 1322827bd09bSSatish Balay 1323827bd09bSSatish Balay Input : 1324827bd09bSSatish Balay Output: 1325827bd09bSSatish Balay Return: 1326827bd09bSSatish Balay Description: 1327827bd09bSSatish Balay ******************************************************************************/ 13280924e98cSBarry Smith PetscErrorCode gs_gop( gs_id *gs, PetscScalar *vals, const char *op) 1329827bd09bSSatish Balay { 1330*d1528f56SBarry Smith PetscErrorCode ierr; 13317b1ae94cSBarry Smith 1332*d1528f56SBarry Smith PetscFunctionBegin; 1333827bd09bSSatish Balay switch (*op) { 1334827bd09bSSatish Balay case '+': 1335827bd09bSSatish Balay gs_gop_plus(gs,vals); 1336827bd09bSSatish Balay break; 1337827bd09bSSatish Balay case '*': 1338827bd09bSSatish Balay gs_gop_times(gs,vals); 1339827bd09bSSatish Balay break; 1340827bd09bSSatish Balay case 'a': 1341827bd09bSSatish Balay gs_gop_min_abs(gs,vals); 1342827bd09bSSatish Balay break; 1343827bd09bSSatish Balay case 'A': 1344827bd09bSSatish Balay gs_gop_max_abs(gs,vals); 1345827bd09bSSatish Balay break; 1346827bd09bSSatish Balay case 'e': 1347827bd09bSSatish Balay gs_gop_exists(gs,vals); 1348827bd09bSSatish Balay break; 1349827bd09bSSatish Balay case 'm': 1350827bd09bSSatish Balay gs_gop_min(gs,vals); 1351827bd09bSSatish Balay break; 1352827bd09bSSatish Balay case 'M': 1353827bd09bSSatish Balay gs_gop_max(gs,vals); break; 1354827bd09bSSatish Balay default: 1355*d1528f56SBarry Smith ierr = PetscInfo1(0,"gs_gop() :: %c is not a valid op",op[0]); 1356*d1528f56SBarry Smith ierr = PetscInfo(0,"gs_gop() :: default :: plus"); 1357827bd09bSSatish Balay gs_gop_plus(gs,vals); 1358827bd09bSSatish Balay break; 1359827bd09bSSatish Balay } 13603fdc5746SBarry Smith PetscFunctionReturn(0); 1361827bd09bSSatish Balay } 1362827bd09bSSatish Balay 1363827bd09bSSatish Balay 1364827bd09bSSatish Balay /****************************************************************************** 1365827bd09bSSatish Balay Function: gather_scatter 1366827bd09bSSatish Balay 1367827bd09bSSatish Balay Input : 1368827bd09bSSatish Balay Output: 1369827bd09bSSatish Balay Return: 1370827bd09bSSatish Balay Description: 1371827bd09bSSatish Balay ******************************************************************************/ 13720924e98cSBarry Smith static PetscErrorCode gs_gop_exists( gs_id *gs, PetscScalar *vals) 1373827bd09bSSatish Balay { 13743fdc5746SBarry Smith PetscFunctionBegin; 1375827bd09bSSatish Balay /* local only operations!!! */ 1376827bd09bSSatish Balay if (gs->num_local) 1377827bd09bSSatish Balay {gs_gop_local_exists(gs,vals);} 1378827bd09bSSatish Balay 1379827bd09bSSatish Balay /* if intersection tree/pairwise and local isn't empty */ 1380827bd09bSSatish Balay if (gs->num_local_gop) 1381827bd09bSSatish Balay { 1382827bd09bSSatish Balay gs_gop_local_in_exists(gs,vals); 1383827bd09bSSatish Balay 1384827bd09bSSatish Balay /* pairwise */ 1385827bd09bSSatish Balay if (gs->num_pairs) 1386827bd09bSSatish Balay {gs_gop_pairwise_exists(gs,vals);} 1387827bd09bSSatish Balay 1388827bd09bSSatish Balay /* tree */ 1389827bd09bSSatish Balay else if (gs->max_left_over) 1390827bd09bSSatish Balay {gs_gop_tree_exists(gs,vals);} 1391827bd09bSSatish Balay 1392827bd09bSSatish Balay gs_gop_local_out(gs,vals); 1393827bd09bSSatish Balay } 1394827bd09bSSatish Balay /* if intersection tree/pairwise and local is empty */ 1395827bd09bSSatish Balay else 1396827bd09bSSatish Balay { 1397827bd09bSSatish Balay /* pairwise */ 1398827bd09bSSatish Balay if (gs->num_pairs) 1399827bd09bSSatish Balay {gs_gop_pairwise_exists(gs,vals);} 1400827bd09bSSatish Balay 1401827bd09bSSatish Balay /* tree */ 1402827bd09bSSatish Balay else if (gs->max_left_over) 1403827bd09bSSatish Balay {gs_gop_tree_exists(gs,vals);} 1404827bd09bSSatish Balay } 14053fdc5746SBarry Smith PetscFunctionReturn(0); 1406827bd09bSSatish Balay } 1407827bd09bSSatish Balay 1408827bd09bSSatish Balay 1409827bd09bSSatish Balay 1410827bd09bSSatish Balay /****************************************************************************** 1411827bd09bSSatish Balay Function: gather_scatter 1412827bd09bSSatish Balay 1413827bd09bSSatish Balay Input : 1414827bd09bSSatish Balay Output: 1415827bd09bSSatish Balay Return: 1416827bd09bSSatish Balay Description: 1417827bd09bSSatish Balay ******************************************************************************/ 14180924e98cSBarry Smith static PetscErrorCode gs_gop_local_exists( gs_id *gs, PetscScalar *vals) 1419827bd09bSSatish Balay { 142052f87cdaSBarry Smith PetscInt *num, *map, **reduce; 1421a501084fSBarry Smith PetscScalar tmp; 1422827bd09bSSatish Balay 14233fdc5746SBarry Smith PetscFunctionBegin; 1424827bd09bSSatish Balay num = gs->num_local_reduce; 1425827bd09bSSatish Balay reduce = gs->local_reduce; 1426827bd09bSSatish Balay while ((map = *reduce)) 1427827bd09bSSatish Balay { 1428827bd09bSSatish Balay num ++; 1429827bd09bSSatish Balay tmp = 0.0; 1430827bd09bSSatish Balay while (*map >= 0) 1431827bd09bSSatish Balay {tmp = EXISTS(tmp,*(vals + *map)); map++;} 1432827bd09bSSatish Balay 1433827bd09bSSatish Balay map = *reduce++; 1434827bd09bSSatish Balay while (*map >= 0) 1435827bd09bSSatish Balay {*(vals + *map++) = tmp;} 1436827bd09bSSatish Balay } 14373fdc5746SBarry Smith PetscFunctionReturn(0); 1438827bd09bSSatish Balay } 1439827bd09bSSatish Balay 14407b1ae94cSBarry Smith /******************************************************************************/ 14410924e98cSBarry Smith static PetscErrorCode gs_gop_local_in_exists( gs_id *gs, PetscScalar *vals) 1442827bd09bSSatish Balay { 144352f87cdaSBarry Smith PetscInt *num, *map, **reduce; 1444a501084fSBarry Smith PetscScalar *base; 1445827bd09bSSatish Balay 14463fdc5746SBarry Smith PetscFunctionBegin; 1447827bd09bSSatish Balay num = gs->num_gop_local_reduce; 1448827bd09bSSatish Balay reduce = gs->gop_local_reduce; 1449827bd09bSSatish Balay while ((map = *reduce++)) 1450827bd09bSSatish Balay { 1451827bd09bSSatish Balay num++; 1452827bd09bSSatish Balay base = vals + *map++; 1453827bd09bSSatish Balay while (*map >= 0) 1454827bd09bSSatish Balay {*base = EXISTS(*base,*(vals + *map)); map++;} 1455827bd09bSSatish Balay } 14563fdc5746SBarry Smith PetscFunctionReturn(0); 1457827bd09bSSatish Balay } 1458827bd09bSSatish Balay 14590924e98cSBarry Smith static PetscErrorCode gs_gop_pairwise_exists( gs_id *gs, PetscScalar *in_vals) 1460827bd09bSSatish Balay { 1461a501084fSBarry Smith PetscScalar *dptr1, *dptr2, *dptr3, *in1, *in2; 146252f87cdaSBarry Smith PetscInt *iptr, *msg_list, *msg_size, **msg_nodes; 146352f87cdaSBarry Smith PetscInt *pw, *list, *size, **nodes; 1464827bd09bSSatish Balay MPI_Request *msg_ids_in, *msg_ids_out, *ids_in, *ids_out; 1465827bd09bSSatish Balay MPI_Status status; 14663fdc5746SBarry Smith PetscErrorCode ierr; 1467827bd09bSSatish Balay 14683fdc5746SBarry Smith PetscFunctionBegin; 1469a501084fSBarry Smith /* strip and load s */ 1470827bd09bSSatish Balay msg_list =list = gs->pair_list; 1471827bd09bSSatish Balay msg_size =size = gs->msg_sizes; 1472827bd09bSSatish Balay msg_nodes=nodes = gs->node_list; 1473827bd09bSSatish Balay iptr=pw = gs->pw_elm_list; 1474827bd09bSSatish Balay dptr1=dptr3 = gs->pw_vals; 1475827bd09bSSatish Balay msg_ids_in = ids_in = gs->msg_ids_in; 1476827bd09bSSatish Balay msg_ids_out = ids_out = gs->msg_ids_out; 1477827bd09bSSatish Balay dptr2 = gs->out; 1478827bd09bSSatish Balay in1=in2 = gs->in; 1479827bd09bSSatish Balay 1480827bd09bSSatish Balay /* post the receives */ 1481827bd09bSSatish Balay do 1482827bd09bSSatish Balay { 1483827bd09bSSatish Balay /* Should MPI_ANY_SOURCE be replaced by *list ? In that case do the 1484827bd09bSSatish Balay second one *list and do list++ afterwards */ 14853fdc5746SBarry Smith ierr = MPI_Irecv(in1, *size, MPIU_SCALAR, MPI_ANY_SOURCE, MSGTAG1 + *list++, gs->gs_comm, msg_ids_in++);CHKERRQ(ierr); 1486827bd09bSSatish Balay in1 += *size++; 1487827bd09bSSatish Balay } 1488827bd09bSSatish Balay while (*++msg_nodes); 1489827bd09bSSatish Balay msg_nodes=nodes; 1490827bd09bSSatish Balay 1491827bd09bSSatish Balay /* load gs values into in out gs buffers */ 1492827bd09bSSatish Balay while (*iptr >= 0) 1493827bd09bSSatish Balay {*dptr3++ = *(in_vals + *iptr++);} 1494827bd09bSSatish Balay 1495827bd09bSSatish Balay /* load out buffers and post the sends */ 1496827bd09bSSatish Balay while ((iptr = *msg_nodes++)) 1497827bd09bSSatish Balay { 1498827bd09bSSatish Balay dptr3 = dptr2; 1499827bd09bSSatish Balay while (*iptr >= 0) 1500827bd09bSSatish Balay {*dptr2++ = *(dptr1 + *iptr++);} 1501827bd09bSSatish Balay /* CHECK PERSISTENT COMMS MODE FOR ALL THIS STUFF */ 1502827bd09bSSatish Balay /* is msg_ids_out++ correct? */ 15033fdc5746SBarry Smith ierr = MPI_Isend(dptr3, *msg_size++, MPIU_SCALAR, *msg_list++, MSGTAG1+my_id, gs->gs_comm, msg_ids_out++);CHKERRQ(ierr); 1504827bd09bSSatish Balay } 1505827bd09bSSatish Balay 1506827bd09bSSatish Balay if (gs->max_left_over) 1507827bd09bSSatish Balay {gs_gop_tree_exists(gs,in_vals);} 1508827bd09bSSatish Balay 1509827bd09bSSatish Balay /* process the received data */ 1510827bd09bSSatish Balay msg_nodes=nodes; 1511827bd09bSSatish Balay while ((iptr = *nodes++)) 1512827bd09bSSatish Balay { 1513827bd09bSSatish Balay /* Should I check the return value of MPI_Wait() or status? */ 1514827bd09bSSatish Balay /* Can this loop be replaced by a call to MPI_Waitall()? */ 15153fdc5746SBarry Smith ierr = MPI_Wait(ids_in++, &status);CHKERRQ(ierr); 1516827bd09bSSatish Balay while (*iptr >= 0) 1517827bd09bSSatish Balay {*(dptr1 + *iptr) = EXISTS(*(dptr1 + *iptr),*in2); iptr++; in2++;} 1518827bd09bSSatish Balay } 1519827bd09bSSatish Balay 1520827bd09bSSatish Balay /* replace vals */ 1521827bd09bSSatish Balay while (*pw >= 0) 1522827bd09bSSatish Balay {*(in_vals + *pw++) = *dptr1++;} 1523827bd09bSSatish Balay 1524827bd09bSSatish Balay /* clear isend message handles */ 1525827bd09bSSatish Balay /* This changed for clarity though it could be the same */ 1526827bd09bSSatish Balay while (*msg_nodes++) 1527827bd09bSSatish Balay /* Should I check the return value of MPI_Wait() or status? */ 1528827bd09bSSatish Balay /* Can this loop be replaced by a call to MPI_Waitall()? */ 15293fdc5746SBarry Smith {ierr = MPI_Wait(ids_out++, &status);CHKERRQ(ierr);} 15303fdc5746SBarry Smith PetscFunctionReturn(0); 1531827bd09bSSatish Balay } 15327b1ae94cSBarry Smith /******************************************************************************/ 15330924e98cSBarry Smith static PetscErrorCode gs_gop_tree_exists(gs_id *gs, PetscScalar *vals) 1534827bd09bSSatish Balay { 153552f87cdaSBarry Smith PetscInt size; 153652f87cdaSBarry Smith PetscInt *in, *out; 1537a501084fSBarry Smith PetscScalar *buf, *work; 153852f87cdaSBarry Smith PetscInt op[] = {GL_EXISTS,0}; 1539827bd09bSSatish Balay 15403fdc5746SBarry Smith PetscFunctionBegin; 1541827bd09bSSatish Balay in = gs->tree_map_in; 1542827bd09bSSatish Balay out = gs->tree_map_out; 1543827bd09bSSatish Balay buf = gs->tree_buf; 1544827bd09bSSatish Balay work = gs->tree_work; 1545827bd09bSSatish Balay size = gs->tree_nel; 1546827bd09bSSatish Balay 1547827bd09bSSatish Balay rvec_zero(buf,size); 1548827bd09bSSatish Balay 1549827bd09bSSatish Balay while (*in >= 0) 1550827bd09bSSatish Balay { 1551827bd09bSSatish Balay /* 1552827bd09bSSatish Balay printf("%d :: out=%d\n",my_id,*out); 1553827bd09bSSatish Balay printf("%d :: in=%d\n",my_id,*in); 1554827bd09bSSatish Balay */ 1555827bd09bSSatish Balay *(buf + *out++) = *(vals + *in++); 1556827bd09bSSatish Balay } 1557827bd09bSSatish Balay 1558827bd09bSSatish Balay grop(buf,work,size,op); 1559827bd09bSSatish Balay 1560827bd09bSSatish Balay in = gs->tree_map_in; 1561827bd09bSSatish Balay out = gs->tree_map_out; 1562827bd09bSSatish Balay 1563827bd09bSSatish Balay while (*in >= 0) 1564827bd09bSSatish Balay {*(vals + *in++) = *(buf + *out++);} 15653fdc5746SBarry Smith PetscFunctionReturn(0); 1566827bd09bSSatish Balay } 1567827bd09bSSatish Balay 15687b1ae94cSBarry Smith /*******************************************************************************/ 15690924e98cSBarry Smith static PetscErrorCode gs_gop_max_abs( gs_id *gs, PetscScalar *vals) 1570827bd09bSSatish Balay { 15713fdc5746SBarry Smith PetscFunctionBegin; 1572827bd09bSSatish Balay /* local only operations!!! */ 1573827bd09bSSatish Balay if (gs->num_local) 1574827bd09bSSatish Balay {gs_gop_local_max_abs(gs,vals);} 1575827bd09bSSatish Balay 1576827bd09bSSatish Balay /* if intersection tree/pairwise and local isn't empty */ 1577827bd09bSSatish Balay if (gs->num_local_gop) 1578827bd09bSSatish Balay { 1579827bd09bSSatish Balay gs_gop_local_in_max_abs(gs,vals); 1580827bd09bSSatish Balay 1581827bd09bSSatish Balay /* pairwise */ 1582827bd09bSSatish Balay if (gs->num_pairs) 1583827bd09bSSatish Balay {gs_gop_pairwise_max_abs(gs,vals);} 1584827bd09bSSatish Balay 1585827bd09bSSatish Balay /* tree */ 1586827bd09bSSatish Balay else if (gs->max_left_over) 1587827bd09bSSatish Balay {gs_gop_tree_max_abs(gs,vals);} 1588827bd09bSSatish Balay 1589827bd09bSSatish Balay gs_gop_local_out(gs,vals); 1590827bd09bSSatish Balay } 1591827bd09bSSatish Balay /* if intersection tree/pairwise and local is empty */ 1592827bd09bSSatish Balay else 1593827bd09bSSatish Balay { 1594827bd09bSSatish Balay /* pairwise */ 1595827bd09bSSatish Balay if (gs->num_pairs) 1596827bd09bSSatish Balay {gs_gop_pairwise_max_abs(gs,vals);} 1597827bd09bSSatish Balay 1598827bd09bSSatish Balay /* tree */ 1599827bd09bSSatish Balay else if (gs->max_left_over) 1600827bd09bSSatish Balay {gs_gop_tree_max_abs(gs,vals);} 1601827bd09bSSatish Balay } 16023fdc5746SBarry Smith PetscFunctionReturn(0); 1603827bd09bSSatish Balay } 1604827bd09bSSatish Balay 16057b1ae94cSBarry Smith /******************************************************************************/ 16060924e98cSBarry Smith static PetscErrorCode gs_gop_local_max_abs( gs_id *gs, PetscScalar *vals) 1607827bd09bSSatish Balay { 160852f87cdaSBarry Smith PetscInt *num, *map, **reduce; 1609a501084fSBarry Smith PetscScalar tmp; 1610827bd09bSSatish Balay 16113fdc5746SBarry Smith PetscFunctionBegin; 1612827bd09bSSatish Balay num = gs->num_local_reduce; 1613827bd09bSSatish Balay reduce = gs->local_reduce; 1614827bd09bSSatish Balay while ((map = *reduce)) 1615827bd09bSSatish Balay { 1616827bd09bSSatish Balay num ++; 1617827bd09bSSatish Balay tmp = 0.0; 1618827bd09bSSatish Balay while (*map >= 0) 1619827bd09bSSatish Balay {tmp = MAX_FABS(tmp,*(vals + *map)); map++;} 1620827bd09bSSatish Balay 1621827bd09bSSatish Balay map = *reduce++; 1622827bd09bSSatish Balay while (*map >= 0) 1623827bd09bSSatish Balay {*(vals + *map++) = tmp;} 1624827bd09bSSatish Balay } 16253fdc5746SBarry Smith PetscFunctionReturn(0); 1626827bd09bSSatish Balay } 1627827bd09bSSatish Balay 16287b1ae94cSBarry Smith /******************************************************************************/ 16290924e98cSBarry Smith static PetscErrorCode gs_gop_local_in_max_abs( gs_id *gs, PetscScalar *vals) 1630827bd09bSSatish Balay { 163152f87cdaSBarry Smith PetscInt *num, *map, **reduce; 1632a501084fSBarry Smith PetscScalar *base; 1633827bd09bSSatish Balay 16343fdc5746SBarry Smith PetscFunctionBegin; 1635827bd09bSSatish Balay num = gs->num_gop_local_reduce; 1636827bd09bSSatish Balay reduce = gs->gop_local_reduce; 1637827bd09bSSatish Balay while ((map = *reduce++)) 1638827bd09bSSatish Balay { 1639827bd09bSSatish Balay num++; 1640827bd09bSSatish Balay base = vals + *map++; 1641827bd09bSSatish Balay while (*map >= 0) 1642827bd09bSSatish Balay {*base = MAX_FABS(*base,*(vals + *map)); map++;} 1643827bd09bSSatish Balay } 16443fdc5746SBarry Smith PetscFunctionReturn(0); 1645827bd09bSSatish Balay } 1646827bd09bSSatish Balay 16477b1ae94cSBarry Smith /******************************************************************************/ 16480924e98cSBarry Smith static PetscErrorCode gs_gop_pairwise_max_abs( gs_id *gs, PetscScalar *in_vals) 1649827bd09bSSatish Balay { 1650a501084fSBarry Smith PetscScalar *dptr1, *dptr2, *dptr3, *in1, *in2; 165152f87cdaSBarry Smith PetscInt *iptr, *msg_list, *msg_size, **msg_nodes; 165252f87cdaSBarry Smith PetscInt *pw, *list, *size, **nodes; 1653827bd09bSSatish Balay MPI_Request *msg_ids_in, *msg_ids_out, *ids_in, *ids_out; 1654827bd09bSSatish Balay MPI_Status status; 16553fdc5746SBarry Smith PetscErrorCode ierr; 1656827bd09bSSatish Balay 16573fdc5746SBarry Smith PetscFunctionBegin; 1658a501084fSBarry Smith /* strip and load s */ 1659827bd09bSSatish Balay msg_list =list = gs->pair_list; 1660827bd09bSSatish Balay msg_size =size = gs->msg_sizes; 1661827bd09bSSatish Balay msg_nodes=nodes = gs->node_list; 1662827bd09bSSatish Balay iptr=pw = gs->pw_elm_list; 1663827bd09bSSatish Balay dptr1=dptr3 = gs->pw_vals; 1664827bd09bSSatish Balay msg_ids_in = ids_in = gs->msg_ids_in; 1665827bd09bSSatish Balay msg_ids_out = ids_out = gs->msg_ids_out; 1666827bd09bSSatish Balay dptr2 = gs->out; 1667827bd09bSSatish Balay in1=in2 = gs->in; 1668827bd09bSSatish Balay 1669827bd09bSSatish Balay /* post the receives */ 1670827bd09bSSatish Balay /* msg_nodes=nodes; */ 1671827bd09bSSatish Balay do 1672827bd09bSSatish Balay { 1673827bd09bSSatish Balay /* Should MPI_ANY_SOURCE be replaced by *list ? In that case do the 1674827bd09bSSatish Balay second one *list and do list++ afterwards */ 16753fdc5746SBarry Smith ierr = MPI_Irecv(in1, *size, MPIU_SCALAR, MPI_ANY_SOURCE, MSGTAG1 + *list++, gs->gs_comm, msg_ids_in++);CHKERRQ(ierr); 1676827bd09bSSatish Balay in1 += *size++; 1677827bd09bSSatish Balay } 1678827bd09bSSatish Balay while (*++msg_nodes); 1679827bd09bSSatish Balay msg_nodes=nodes; 1680827bd09bSSatish Balay 1681827bd09bSSatish Balay /* load gs values into in out gs buffers */ 1682827bd09bSSatish Balay while (*iptr >= 0) 1683827bd09bSSatish Balay {*dptr3++ = *(in_vals + *iptr++);} 1684827bd09bSSatish Balay 1685827bd09bSSatish Balay /* load out buffers and post the sends */ 1686827bd09bSSatish Balay while ((iptr = *msg_nodes++)) 1687827bd09bSSatish Balay { 1688827bd09bSSatish Balay dptr3 = dptr2; 1689827bd09bSSatish Balay while (*iptr >= 0) 1690827bd09bSSatish Balay {*dptr2++ = *(dptr1 + *iptr++);} 1691827bd09bSSatish Balay /* CHECK PERSISTENT COMMS MODE FOR ALL THIS STUFF */ 1692827bd09bSSatish Balay /* is msg_ids_out++ correct? */ 16933fdc5746SBarry Smith ierr = MPI_Isend(dptr3, *msg_size++, MPIU_SCALAR, *msg_list++, MSGTAG1+my_id, gs->gs_comm, msg_ids_out++);CHKERRQ(ierr); 1694827bd09bSSatish Balay } 1695827bd09bSSatish Balay 1696827bd09bSSatish Balay if (gs->max_left_over) 1697827bd09bSSatish Balay {gs_gop_tree_max_abs(gs,in_vals);} 1698827bd09bSSatish Balay 1699827bd09bSSatish Balay /* process the received data */ 1700827bd09bSSatish Balay msg_nodes=nodes; 1701827bd09bSSatish Balay while ((iptr = *nodes++)) 1702827bd09bSSatish Balay { 1703827bd09bSSatish Balay /* Should I check the return value of MPI_Wait() or status? */ 1704827bd09bSSatish Balay /* Can this loop be replaced by a call to MPI_Waitall()? */ 17053fdc5746SBarry Smith ierr = MPI_Wait(ids_in++, &status);CHKERRQ(ierr); 1706827bd09bSSatish Balay while (*iptr >= 0) 1707827bd09bSSatish Balay {*(dptr1 + *iptr) = MAX_FABS(*(dptr1 + *iptr),*in2); iptr++; in2++;} 1708827bd09bSSatish Balay } 1709827bd09bSSatish Balay 1710827bd09bSSatish Balay /* replace vals */ 1711827bd09bSSatish Balay while (*pw >= 0) 1712827bd09bSSatish Balay {*(in_vals + *pw++) = *dptr1++;} 1713827bd09bSSatish Balay 1714827bd09bSSatish Balay /* clear isend message handles */ 1715827bd09bSSatish Balay /* This changed for clarity though it could be the same */ 1716827bd09bSSatish Balay while (*msg_nodes++) 1717827bd09bSSatish Balay /* Should I check the return value of MPI_Wait() or status? */ 1718827bd09bSSatish Balay /* Can this loop be replaced by a call to MPI_Waitall()? */ 17193fdc5746SBarry Smith {ierr = MPI_Wait(ids_out++, &status);CHKERRQ(ierr);} 17203fdc5746SBarry Smith PetscFunctionReturn(0); 1721827bd09bSSatish Balay } 1722827bd09bSSatish Balay 17237b1ae94cSBarry Smith /******************************************************************************/ 17240924e98cSBarry Smith static PetscErrorCode gs_gop_tree_max_abs(gs_id *gs, PetscScalar *vals) 1725827bd09bSSatish Balay { 172652f87cdaSBarry Smith PetscInt size; 172752f87cdaSBarry Smith PetscInt *in, *out; 1728a501084fSBarry Smith PetscScalar *buf, *work; 172952f87cdaSBarry Smith PetscInt op[] = {GL_MAX_ABS,0}; 1730827bd09bSSatish Balay 17313fdc5746SBarry Smith PetscFunctionBegin; 1732827bd09bSSatish Balay in = gs->tree_map_in; 1733827bd09bSSatish Balay out = gs->tree_map_out; 1734827bd09bSSatish Balay buf = gs->tree_buf; 1735827bd09bSSatish Balay work = gs->tree_work; 1736827bd09bSSatish Balay size = gs->tree_nel; 1737827bd09bSSatish Balay 1738827bd09bSSatish Balay rvec_zero(buf,size); 1739827bd09bSSatish Balay 1740827bd09bSSatish Balay while (*in >= 0) 1741827bd09bSSatish Balay { 1742827bd09bSSatish Balay /* 1743827bd09bSSatish Balay printf("%d :: out=%d\n",my_id,*out); 1744827bd09bSSatish Balay printf("%d :: in=%d\n",my_id,*in); 1745827bd09bSSatish Balay */ 1746827bd09bSSatish Balay *(buf + *out++) = *(vals + *in++); 1747827bd09bSSatish Balay } 1748827bd09bSSatish Balay 1749827bd09bSSatish Balay grop(buf,work,size,op); 1750827bd09bSSatish Balay 1751827bd09bSSatish Balay in = gs->tree_map_in; 1752827bd09bSSatish Balay out = gs->tree_map_out; 1753827bd09bSSatish Balay 1754827bd09bSSatish Balay while (*in >= 0) 1755827bd09bSSatish Balay {*(vals + *in++) = *(buf + *out++);} 17563fdc5746SBarry Smith PetscFunctionReturn(0); 1757827bd09bSSatish Balay } 1758827bd09bSSatish Balay 17597b1ae94cSBarry Smith /******************************************************************************/ 17600924e98cSBarry Smith static PetscErrorCode gs_gop_max( gs_id *gs, PetscScalar *vals) 1761827bd09bSSatish Balay { 17623fdc5746SBarry Smith PetscFunctionBegin; 1763827bd09bSSatish Balay /* local only operations!!! */ 1764827bd09bSSatish Balay if (gs->num_local) 1765827bd09bSSatish Balay {gs_gop_local_max(gs,vals);} 1766827bd09bSSatish Balay 1767827bd09bSSatish Balay /* if intersection tree/pairwise and local isn't empty */ 1768827bd09bSSatish Balay if (gs->num_local_gop) 1769827bd09bSSatish Balay { 1770827bd09bSSatish Balay gs_gop_local_in_max(gs,vals); 1771827bd09bSSatish Balay 1772827bd09bSSatish Balay /* pairwise */ 1773827bd09bSSatish Balay if (gs->num_pairs) 1774827bd09bSSatish Balay {gs_gop_pairwise_max(gs,vals);} 1775827bd09bSSatish Balay 1776827bd09bSSatish Balay /* tree */ 1777827bd09bSSatish Balay else if (gs->max_left_over) 1778827bd09bSSatish Balay {gs_gop_tree_max(gs,vals);} 1779827bd09bSSatish Balay 1780827bd09bSSatish Balay gs_gop_local_out(gs,vals); 1781827bd09bSSatish Balay } 1782827bd09bSSatish Balay /* if intersection tree/pairwise and local is empty */ 1783827bd09bSSatish Balay else 1784827bd09bSSatish Balay { 1785827bd09bSSatish Balay /* pairwise */ 1786827bd09bSSatish Balay if (gs->num_pairs) 1787827bd09bSSatish Balay {gs_gop_pairwise_max(gs,vals);} 1788827bd09bSSatish Balay 1789827bd09bSSatish Balay /* tree */ 1790827bd09bSSatish Balay else if (gs->max_left_over) 1791827bd09bSSatish Balay {gs_gop_tree_max(gs,vals);} 1792827bd09bSSatish Balay } 17933fdc5746SBarry Smith PetscFunctionReturn(0); 1794827bd09bSSatish Balay } 1795827bd09bSSatish Balay 17967b1ae94cSBarry Smith /******************************************************************************/ 17970924e98cSBarry Smith static PetscErrorCode gs_gop_local_max( gs_id *gs, PetscScalar *vals) 1798827bd09bSSatish Balay { 179952f87cdaSBarry Smith PetscInt *num, *map, **reduce; 1800a501084fSBarry Smith PetscScalar tmp; 1801827bd09bSSatish Balay 18023fdc5746SBarry Smith PetscFunctionBegin; 1803827bd09bSSatish Balay num = gs->num_local_reduce; 1804827bd09bSSatish Balay reduce = gs->local_reduce; 1805827bd09bSSatish Balay while ((map = *reduce)) 1806827bd09bSSatish Balay { 1807827bd09bSSatish Balay num ++; 1808827bd09bSSatish Balay tmp = -REAL_MAX; 1809827bd09bSSatish Balay while (*map >= 0) 181039945688SSatish Balay {tmp = PetscMax(tmp,*(vals + *map)); map++;} 1811827bd09bSSatish Balay 1812827bd09bSSatish Balay map = *reduce++; 1813827bd09bSSatish Balay while (*map >= 0) 1814827bd09bSSatish Balay {*(vals + *map++) = tmp;} 1815827bd09bSSatish Balay } 18163fdc5746SBarry Smith PetscFunctionReturn(0); 1817827bd09bSSatish Balay } 1818827bd09bSSatish Balay 18197b1ae94cSBarry Smith /******************************************************************************/ 18200924e98cSBarry Smith static PetscErrorCode gs_gop_local_in_max( gs_id *gs, PetscScalar *vals) 1821827bd09bSSatish Balay { 182252f87cdaSBarry Smith PetscInt *num, *map, **reduce; 1823a501084fSBarry Smith PetscScalar *base; 1824827bd09bSSatish Balay 18253fdc5746SBarry Smith PetscFunctionBegin; 1826827bd09bSSatish Balay num = gs->num_gop_local_reduce; 1827827bd09bSSatish Balay reduce = gs->gop_local_reduce; 1828827bd09bSSatish Balay while ((map = *reduce++)) 1829827bd09bSSatish Balay { 1830827bd09bSSatish Balay num++; 1831827bd09bSSatish Balay base = vals + *map++; 1832827bd09bSSatish Balay while (*map >= 0) 183339945688SSatish Balay {*base = PetscMax(*base,*(vals + *map)); map++;} 1834827bd09bSSatish Balay } 18353fdc5746SBarry Smith PetscFunctionReturn(0); 1836827bd09bSSatish Balay } 1837827bd09bSSatish Balay 18387b1ae94cSBarry Smith /******************************************************************************/ 18390924e98cSBarry Smith static PetscErrorCode gs_gop_pairwise_max( gs_id *gs, PetscScalar *in_vals) 1840827bd09bSSatish Balay { 1841a501084fSBarry Smith PetscScalar *dptr1, *dptr2, *dptr3, *in1, *in2; 184252f87cdaSBarry Smith PetscInt *iptr, *msg_list, *msg_size, **msg_nodes; 184352f87cdaSBarry Smith PetscInt *pw, *list, *size, **nodes; 1844827bd09bSSatish Balay MPI_Request *msg_ids_in, *msg_ids_out, *ids_in, *ids_out; 1845827bd09bSSatish Balay MPI_Status status; 18463fdc5746SBarry Smith PetscErrorCode ierr; 1847827bd09bSSatish Balay 18483fdc5746SBarry Smith PetscFunctionBegin; 1849a501084fSBarry Smith /* strip and load s */ 1850827bd09bSSatish Balay msg_list =list = gs->pair_list; 1851827bd09bSSatish Balay msg_size =size = gs->msg_sizes; 1852827bd09bSSatish Balay msg_nodes=nodes = gs->node_list; 1853827bd09bSSatish Balay iptr=pw = gs->pw_elm_list; 1854827bd09bSSatish Balay dptr1=dptr3 = gs->pw_vals; 1855827bd09bSSatish Balay msg_ids_in = ids_in = gs->msg_ids_in; 1856827bd09bSSatish Balay msg_ids_out = ids_out = gs->msg_ids_out; 1857827bd09bSSatish Balay dptr2 = gs->out; 1858827bd09bSSatish Balay in1=in2 = gs->in; 1859827bd09bSSatish Balay 1860827bd09bSSatish Balay /* post the receives */ 1861827bd09bSSatish Balay /* msg_nodes=nodes; */ 1862827bd09bSSatish Balay do 1863827bd09bSSatish Balay { 1864827bd09bSSatish Balay /* Should MPI_ANY_SOURCE be replaced by *list ? In that case do the 1865827bd09bSSatish Balay second one *list and do list++ afterwards */ 18663fdc5746SBarry Smith ierr = MPI_Irecv(in1, *size, MPIU_SCALAR, MPI_ANY_SOURCE, MSGTAG1 + *list++, gs->gs_comm, msg_ids_in++);CHKERRQ(ierr); 1867827bd09bSSatish Balay in1 += *size++; 1868827bd09bSSatish Balay } 1869827bd09bSSatish Balay while (*++msg_nodes); 1870827bd09bSSatish Balay msg_nodes=nodes; 1871827bd09bSSatish Balay 1872827bd09bSSatish Balay /* load gs values into in out gs buffers */ 1873827bd09bSSatish Balay while (*iptr >= 0) 1874827bd09bSSatish Balay {*dptr3++ = *(in_vals + *iptr++);} 1875827bd09bSSatish Balay 1876827bd09bSSatish Balay /* load out buffers and post the sends */ 1877827bd09bSSatish Balay while ((iptr = *msg_nodes++)) 1878827bd09bSSatish Balay { 1879827bd09bSSatish Balay dptr3 = dptr2; 1880827bd09bSSatish Balay while (*iptr >= 0) 1881827bd09bSSatish Balay {*dptr2++ = *(dptr1 + *iptr++);} 1882827bd09bSSatish Balay /* CHECK PERSISTENT COMMS MODE FOR ALL THIS STUFF */ 1883827bd09bSSatish Balay /* is msg_ids_out++ correct? */ 18843fdc5746SBarry Smith ierr = MPI_Isend(dptr3, *msg_size++, MPIU_SCALAR, *msg_list++, MSGTAG1+my_id, gs->gs_comm, msg_ids_out++);CHKERRQ(ierr); 1885827bd09bSSatish Balay } 1886827bd09bSSatish Balay 1887827bd09bSSatish Balay if (gs->max_left_over) 1888827bd09bSSatish Balay {gs_gop_tree_max(gs,in_vals);} 1889827bd09bSSatish Balay 1890827bd09bSSatish Balay /* process the received data */ 1891827bd09bSSatish Balay msg_nodes=nodes; 1892827bd09bSSatish Balay while ((iptr = *nodes++)) 1893827bd09bSSatish Balay { 1894827bd09bSSatish Balay /* Should I check the return value of MPI_Wait() or status? */ 1895827bd09bSSatish Balay /* Can this loop be replaced by a call to MPI_Waitall()? */ 18963fdc5746SBarry Smith ierr = MPI_Wait(ids_in++, &status);CHKERRQ(ierr); 1897827bd09bSSatish Balay while (*iptr >= 0) 189839945688SSatish Balay {*(dptr1 + *iptr) = PetscMax(*(dptr1 + *iptr),*in2); iptr++; in2++;} 1899827bd09bSSatish Balay } 1900827bd09bSSatish Balay 1901827bd09bSSatish Balay /* replace vals */ 1902827bd09bSSatish Balay while (*pw >= 0) 1903827bd09bSSatish Balay {*(in_vals + *pw++) = *dptr1++;} 1904827bd09bSSatish Balay 1905827bd09bSSatish Balay /* clear isend message handles */ 1906827bd09bSSatish Balay /* This changed for clarity though it could be the same */ 1907827bd09bSSatish Balay while (*msg_nodes++) 1908827bd09bSSatish Balay /* Should I check the return value of MPI_Wait() or status? */ 1909827bd09bSSatish Balay /* Can this loop be replaced by a call to MPI_Waitall()? */ 19103fdc5746SBarry Smith {ierr = MPI_Wait(ids_out++, &status);CHKERRQ(ierr);} 19113fdc5746SBarry Smith PetscFunctionReturn(0); 1912827bd09bSSatish Balay } 1913827bd09bSSatish Balay 19147b1ae94cSBarry Smith /******************************************************************************/ 19150924e98cSBarry Smith static PetscErrorCode gs_gop_tree_max(gs_id *gs, PetscScalar *vals) 1916827bd09bSSatish Balay { 191752f87cdaSBarry Smith PetscInt size; 191852f87cdaSBarry Smith PetscInt *in, *out; 1919a501084fSBarry Smith PetscScalar *buf, *work; 19203fdc5746SBarry Smith PetscErrorCode ierr; 1921827bd09bSSatish Balay 19223fdc5746SBarry Smith PetscFunctionBegin; 1923827bd09bSSatish Balay in = gs->tree_map_in; 1924827bd09bSSatish Balay out = gs->tree_map_out; 1925827bd09bSSatish Balay buf = gs->tree_buf; 1926827bd09bSSatish Balay work = gs->tree_work; 1927827bd09bSSatish Balay size = gs->tree_nel; 1928827bd09bSSatish Balay 1929827bd09bSSatish Balay rvec_set(buf,-REAL_MAX,size); 1930827bd09bSSatish Balay 1931827bd09bSSatish Balay while (*in >= 0) 1932827bd09bSSatish Balay {*(buf + *out++) = *(vals + *in++);} 1933827bd09bSSatish Balay 1934827bd09bSSatish Balay in = gs->tree_map_in; 1935827bd09bSSatish Balay out = gs->tree_map_out; 19363fdc5746SBarry Smith ierr = MPI_Allreduce(buf,work,size,MPIU_SCALAR,MPI_MAX,gs->gs_comm);CHKERRQ(ierr); 1937827bd09bSSatish Balay while (*in >= 0) 1938827bd09bSSatish Balay {*(vals + *in++) = *(work + *out++);} 19393fdc5746SBarry Smith PetscFunctionReturn(0); 1940827bd09bSSatish Balay } 19417b1ae94cSBarry Smith /******************************************************************************/ 19420924e98cSBarry Smith static PetscErrorCode gs_gop_min_abs( gs_id *gs, PetscScalar *vals) 1943827bd09bSSatish Balay { 19443fdc5746SBarry Smith PetscFunctionBegin; 1945827bd09bSSatish Balay /* local only operations!!! */ 1946827bd09bSSatish Balay if (gs->num_local) 1947827bd09bSSatish Balay {gs_gop_local_min_abs(gs,vals);} 1948827bd09bSSatish Balay 1949827bd09bSSatish Balay /* if intersection tree/pairwise and local isn't empty */ 1950827bd09bSSatish Balay if (gs->num_local_gop) 1951827bd09bSSatish Balay { 1952827bd09bSSatish Balay gs_gop_local_in_min_abs(gs,vals); 1953827bd09bSSatish Balay 1954827bd09bSSatish Balay /* pairwise */ 1955827bd09bSSatish Balay if (gs->num_pairs) 1956827bd09bSSatish Balay {gs_gop_pairwise_min_abs(gs,vals);} 1957827bd09bSSatish Balay 1958827bd09bSSatish Balay /* tree */ 1959827bd09bSSatish Balay else if (gs->max_left_over) 1960827bd09bSSatish Balay {gs_gop_tree_min_abs(gs,vals);} 1961827bd09bSSatish Balay 1962827bd09bSSatish Balay gs_gop_local_out(gs,vals); 1963827bd09bSSatish Balay } 1964827bd09bSSatish Balay /* if intersection tree/pairwise and local is empty */ 1965827bd09bSSatish Balay else 1966827bd09bSSatish Balay { 1967827bd09bSSatish Balay /* pairwise */ 1968827bd09bSSatish Balay if (gs->num_pairs) 1969827bd09bSSatish Balay {gs_gop_pairwise_min_abs(gs,vals);} 1970827bd09bSSatish Balay 1971827bd09bSSatish Balay /* tree */ 1972827bd09bSSatish Balay else if (gs->max_left_over) 1973827bd09bSSatish Balay {gs_gop_tree_min_abs(gs,vals);} 1974827bd09bSSatish Balay } 19753fdc5746SBarry Smith PetscFunctionReturn(0); 1976827bd09bSSatish Balay } 1977827bd09bSSatish Balay 19787b1ae94cSBarry Smith /******************************************************************************/ 19790924e98cSBarry Smith static PetscErrorCode gs_gop_local_min_abs( gs_id *gs, PetscScalar *vals) 1980827bd09bSSatish Balay { 198152f87cdaSBarry Smith PetscInt *num, *map, **reduce; 1982a501084fSBarry Smith PetscScalar tmp; 1983827bd09bSSatish Balay 19843fdc5746SBarry Smith PetscFunctionBegin; 1985827bd09bSSatish Balay num = gs->num_local_reduce; 1986827bd09bSSatish Balay reduce = gs->local_reduce; 1987827bd09bSSatish Balay while ((map = *reduce)) 1988827bd09bSSatish Balay { 1989827bd09bSSatish Balay num ++; 1990827bd09bSSatish Balay tmp = REAL_MAX; 1991827bd09bSSatish Balay while (*map >= 0) 1992827bd09bSSatish Balay {tmp = MIN_FABS(tmp,*(vals + *map)); map++;} 1993827bd09bSSatish Balay 1994827bd09bSSatish Balay map = *reduce++; 1995827bd09bSSatish Balay while (*map >= 0) 1996827bd09bSSatish Balay {*(vals + *map++) = tmp;} 1997827bd09bSSatish Balay } 19983fdc5746SBarry Smith PetscFunctionReturn(0); 1999827bd09bSSatish Balay } 2000827bd09bSSatish Balay 20017b1ae94cSBarry Smith /******************************************************************************/ 20020924e98cSBarry Smith static PetscErrorCode gs_gop_local_in_min_abs( gs_id *gs, PetscScalar *vals) 2003827bd09bSSatish Balay { 200452f87cdaSBarry Smith PetscInt *num, *map, **reduce; 2005a501084fSBarry Smith PetscScalar *base; 2006827bd09bSSatish Balay 20073fdc5746SBarry Smith PetscFunctionBegin; 2008827bd09bSSatish Balay num = gs->num_gop_local_reduce; 2009827bd09bSSatish Balay reduce = gs->gop_local_reduce; 2010827bd09bSSatish Balay while ((map = *reduce++)) 2011827bd09bSSatish Balay { 2012827bd09bSSatish Balay num++; 2013827bd09bSSatish Balay base = vals + *map++; 2014827bd09bSSatish Balay while (*map >= 0) 2015827bd09bSSatish Balay {*base = MIN_FABS(*base,*(vals + *map)); map++;} 2016827bd09bSSatish Balay } 20173fdc5746SBarry Smith PetscFunctionReturn(0); 2018827bd09bSSatish Balay } 2019827bd09bSSatish Balay 20207b1ae94cSBarry Smith /******************************************************************************/ 20210924e98cSBarry Smith static PetscErrorCode gs_gop_pairwise_min_abs( gs_id *gs, PetscScalar *in_vals) 2022827bd09bSSatish Balay { 2023a501084fSBarry Smith PetscScalar *dptr1, *dptr2, *dptr3, *in1, *in2; 202452f87cdaSBarry Smith PetscInt *iptr, *msg_list, *msg_size, **msg_nodes; 202552f87cdaSBarry Smith PetscInt *pw, *list, *size, **nodes; 2026827bd09bSSatish Balay MPI_Request *msg_ids_in, *msg_ids_out, *ids_in, *ids_out; 2027827bd09bSSatish Balay MPI_Status status; 20283fdc5746SBarry Smith PetscErrorCode ierr; 2029827bd09bSSatish Balay 20303fdc5746SBarry Smith PetscFunctionBegin; 2031a501084fSBarry Smith /* strip and load s */ 2032827bd09bSSatish Balay msg_list =list = gs->pair_list; 2033827bd09bSSatish Balay msg_size =size = gs->msg_sizes; 2034827bd09bSSatish Balay msg_nodes=nodes = gs->node_list; 2035827bd09bSSatish Balay iptr=pw = gs->pw_elm_list; 2036827bd09bSSatish Balay dptr1=dptr3 = gs->pw_vals; 2037827bd09bSSatish Balay msg_ids_in = ids_in = gs->msg_ids_in; 2038827bd09bSSatish Balay msg_ids_out = ids_out = gs->msg_ids_out; 2039827bd09bSSatish Balay dptr2 = gs->out; 2040827bd09bSSatish Balay in1=in2 = gs->in; 2041827bd09bSSatish Balay 2042827bd09bSSatish Balay /* post the receives */ 2043827bd09bSSatish Balay /* msg_nodes=nodes; */ 2044827bd09bSSatish Balay do 2045827bd09bSSatish Balay { 2046827bd09bSSatish Balay /* Should MPI_ANY_SOURCE be replaced by *list ? In that case do the 2047827bd09bSSatish Balay second one *list and do list++ afterwards */ 20483fdc5746SBarry Smith ierr = MPI_Irecv(in1, *size, MPIU_SCALAR, MPI_ANY_SOURCE, MSGTAG1 + *list++, gs->gs_comm, msg_ids_in++);CHKERRQ(ierr); 2049827bd09bSSatish Balay in1 += *size++; 2050827bd09bSSatish Balay } 2051827bd09bSSatish Balay while (*++msg_nodes); 2052827bd09bSSatish Balay msg_nodes=nodes; 2053827bd09bSSatish Balay 2054827bd09bSSatish Balay /* load gs values into in out gs buffers */ 2055827bd09bSSatish Balay while (*iptr >= 0) 2056827bd09bSSatish Balay {*dptr3++ = *(in_vals + *iptr++);} 2057827bd09bSSatish Balay 2058827bd09bSSatish Balay /* load out buffers and post the sends */ 2059827bd09bSSatish Balay while ((iptr = *msg_nodes++)) 2060827bd09bSSatish Balay { 2061827bd09bSSatish Balay dptr3 = dptr2; 2062827bd09bSSatish Balay while (*iptr >= 0) 2063827bd09bSSatish Balay {*dptr2++ = *(dptr1 + *iptr++);} 2064827bd09bSSatish Balay /* CHECK PERSISTENT COMMS MODE FOR ALL THIS STUFF */ 2065827bd09bSSatish Balay /* is msg_ids_out++ correct? */ 20663fdc5746SBarry Smith ierr = MPI_Isend(dptr3, *msg_size++, MPIU_SCALAR, *msg_list++, MSGTAG1+my_id, gs->gs_comm, msg_ids_out++);CHKERRQ(ierr); 2067827bd09bSSatish Balay } 2068827bd09bSSatish Balay 2069827bd09bSSatish Balay if (gs->max_left_over) 2070827bd09bSSatish Balay {gs_gop_tree_min_abs(gs,in_vals);} 2071827bd09bSSatish Balay 2072827bd09bSSatish Balay /* process the received data */ 2073827bd09bSSatish Balay msg_nodes=nodes; 2074827bd09bSSatish Balay while ((iptr = *nodes++)) 2075827bd09bSSatish Balay { 2076827bd09bSSatish Balay /* Should I check the return value of MPI_Wait() or status? */ 2077827bd09bSSatish Balay /* Can this loop be replaced by a call to MPI_Waitall()? */ 20783fdc5746SBarry Smith ierr = MPI_Wait(ids_in++, &status);CHKERRQ(ierr); 2079827bd09bSSatish Balay while (*iptr >= 0) 2080827bd09bSSatish Balay {*(dptr1 + *iptr) = MIN_FABS(*(dptr1 + *iptr),*in2); iptr++; in2++;} 2081827bd09bSSatish Balay } 2082827bd09bSSatish Balay 2083827bd09bSSatish Balay /* replace vals */ 2084827bd09bSSatish Balay while (*pw >= 0) 2085827bd09bSSatish Balay {*(in_vals + *pw++) = *dptr1++;} 2086827bd09bSSatish Balay 2087827bd09bSSatish Balay /* clear isend message handles */ 2088827bd09bSSatish Balay /* This changed for clarity though it could be the same */ 2089827bd09bSSatish Balay while (*msg_nodes++) 2090827bd09bSSatish Balay /* Should I check the return value of MPI_Wait() or status? */ 2091827bd09bSSatish Balay /* Can this loop be replaced by a call to MPI_Waitall()? */ 20923fdc5746SBarry Smith {ierr = MPI_Wait(ids_out++, &status);CHKERRQ(ierr);} 20933fdc5746SBarry Smith PetscFunctionReturn(0); 2094827bd09bSSatish Balay } 2095827bd09bSSatish Balay 20967b1ae94cSBarry Smith /******************************************************************************/ 20970924e98cSBarry Smith static PetscErrorCode gs_gop_tree_min_abs(gs_id *gs, PetscScalar *vals) 2098827bd09bSSatish Balay { 209952f87cdaSBarry Smith PetscInt size; 210052f87cdaSBarry Smith PetscInt *in, *out; 2101a501084fSBarry Smith PetscScalar *buf, *work; 210252f87cdaSBarry Smith PetscInt op[] = {GL_MIN_ABS,0}; 2103827bd09bSSatish Balay 21043fdc5746SBarry Smith PetscFunctionBegin; 2105827bd09bSSatish Balay in = gs->tree_map_in; 2106827bd09bSSatish Balay out = gs->tree_map_out; 2107827bd09bSSatish Balay buf = gs->tree_buf; 2108827bd09bSSatish Balay work = gs->tree_work; 2109827bd09bSSatish Balay size = gs->tree_nel; 2110827bd09bSSatish Balay 2111827bd09bSSatish Balay rvec_set(buf,REAL_MAX,size); 2112827bd09bSSatish Balay 2113827bd09bSSatish Balay while (*in >= 0) 2114827bd09bSSatish Balay {*(buf + *out++) = *(vals + *in++);} 2115827bd09bSSatish Balay 2116827bd09bSSatish Balay in = gs->tree_map_in; 2117827bd09bSSatish Balay out = gs->tree_map_out; 2118827bd09bSSatish Balay grop(buf,work,size,op); 2119827bd09bSSatish Balay while (*in >= 0) 2120827bd09bSSatish Balay {*(vals + *in++) = *(buf + *out++);} 21213fdc5746SBarry Smith PetscFunctionReturn(0); 2122827bd09bSSatish Balay } 2123827bd09bSSatish Balay 21247b1ae94cSBarry Smith /******************************************************************************/ 21250924e98cSBarry Smith static PetscErrorCode gs_gop_min( gs_id *gs, PetscScalar *vals) 2126827bd09bSSatish Balay { 21273fdc5746SBarry Smith PetscFunctionBegin; 2128827bd09bSSatish Balay /* local only operations!!! */ 2129827bd09bSSatish Balay if (gs->num_local) 2130827bd09bSSatish Balay {gs_gop_local_min(gs,vals);} 2131827bd09bSSatish Balay 2132827bd09bSSatish Balay /* if intersection tree/pairwise and local isn't empty */ 2133827bd09bSSatish Balay if (gs->num_local_gop) 2134827bd09bSSatish Balay { 2135827bd09bSSatish Balay gs_gop_local_in_min(gs,vals); 2136827bd09bSSatish Balay 2137827bd09bSSatish Balay /* pairwise */ 2138827bd09bSSatish Balay if (gs->num_pairs) 2139827bd09bSSatish Balay {gs_gop_pairwise_min(gs,vals);} 2140827bd09bSSatish Balay 2141827bd09bSSatish Balay /* tree */ 2142827bd09bSSatish Balay else if (gs->max_left_over) 2143827bd09bSSatish Balay {gs_gop_tree_min(gs,vals);} 2144827bd09bSSatish Balay 2145827bd09bSSatish Balay gs_gop_local_out(gs,vals); 2146827bd09bSSatish Balay } 2147827bd09bSSatish Balay /* if intersection tree/pairwise and local is empty */ 2148827bd09bSSatish Balay else 2149827bd09bSSatish Balay { 2150827bd09bSSatish Balay /* pairwise */ 2151827bd09bSSatish Balay if (gs->num_pairs) 2152827bd09bSSatish Balay {gs_gop_pairwise_min(gs,vals);} 2153827bd09bSSatish Balay 2154827bd09bSSatish Balay /* tree */ 2155827bd09bSSatish Balay else if (gs->max_left_over) 2156827bd09bSSatish Balay {gs_gop_tree_min(gs,vals);} 2157827bd09bSSatish Balay } 21583fdc5746SBarry Smith PetscFunctionReturn(0); 2159827bd09bSSatish Balay } 2160827bd09bSSatish Balay 21617b1ae94cSBarry Smith /******************************************************************************/ 21620924e98cSBarry Smith static PetscErrorCode gs_gop_local_min( gs_id *gs, PetscScalar *vals) 2163827bd09bSSatish Balay { 216452f87cdaSBarry Smith PetscInt *num, *map, **reduce; 2165a501084fSBarry Smith PetscScalar tmp; 21663fdc5746SBarry Smith PetscFunctionBegin; 2167827bd09bSSatish Balay num = gs->num_local_reduce; 2168827bd09bSSatish Balay reduce = gs->local_reduce; 2169827bd09bSSatish Balay while ((map = *reduce)) 2170827bd09bSSatish Balay { 2171827bd09bSSatish Balay num ++; 2172827bd09bSSatish Balay tmp = REAL_MAX; 2173827bd09bSSatish Balay while (*map >= 0) 217439945688SSatish Balay {tmp = PetscMin(tmp,*(vals + *map)); map++;} 2175827bd09bSSatish Balay 2176827bd09bSSatish Balay map = *reduce++; 2177827bd09bSSatish Balay while (*map >= 0) 2178827bd09bSSatish Balay {*(vals + *map++) = tmp;} 2179827bd09bSSatish Balay } 21803fdc5746SBarry Smith PetscFunctionReturn(0); 2181827bd09bSSatish Balay } 2182827bd09bSSatish Balay 21837b1ae94cSBarry Smith /******************************************************************************/ 21840924e98cSBarry Smith static PetscErrorCode gs_gop_local_in_min( gs_id *gs, PetscScalar *vals) 2185827bd09bSSatish Balay { 218652f87cdaSBarry Smith PetscInt *num, *map, **reduce; 2187a501084fSBarry Smith PetscScalar *base; 2188827bd09bSSatish Balay 21893fdc5746SBarry Smith PetscFunctionBegin; 2190827bd09bSSatish Balay num = gs->num_gop_local_reduce; 2191827bd09bSSatish Balay reduce = gs->gop_local_reduce; 2192827bd09bSSatish Balay while ((map = *reduce++)) 2193827bd09bSSatish Balay { 2194827bd09bSSatish Balay num++; 2195827bd09bSSatish Balay base = vals + *map++; 2196827bd09bSSatish Balay while (*map >= 0) 219739945688SSatish Balay {*base = PetscMin(*base,*(vals + *map)); map++;} 2198827bd09bSSatish Balay } 21993fdc5746SBarry Smith PetscFunctionReturn(0); 2200827bd09bSSatish Balay } 2201827bd09bSSatish Balay 22027b1ae94cSBarry Smith /******************************************************************************/ 22030924e98cSBarry Smith static PetscErrorCode gs_gop_pairwise_min( gs_id *gs, PetscScalar *in_vals) 2204827bd09bSSatish Balay { 2205a501084fSBarry Smith PetscScalar *dptr1, *dptr2, *dptr3, *in1, *in2; 220652f87cdaSBarry Smith PetscInt *iptr, *msg_list, *msg_size, **msg_nodes; 220752f87cdaSBarry Smith PetscInt *pw, *list, *size, **nodes; 2208827bd09bSSatish Balay MPI_Request *msg_ids_in, *msg_ids_out, *ids_in, *ids_out; 2209827bd09bSSatish Balay MPI_Status status; 22103fdc5746SBarry Smith PetscErrorCode ierr; 2211827bd09bSSatish Balay 22123fdc5746SBarry Smith PetscFunctionBegin; 2213a501084fSBarry Smith /* strip and load s */ 2214827bd09bSSatish Balay msg_list =list = gs->pair_list; 2215827bd09bSSatish Balay msg_size =size = gs->msg_sizes; 2216827bd09bSSatish Balay msg_nodes=nodes = gs->node_list; 2217827bd09bSSatish Balay iptr=pw = gs->pw_elm_list; 2218827bd09bSSatish Balay dptr1=dptr3 = gs->pw_vals; 2219827bd09bSSatish Balay msg_ids_in = ids_in = gs->msg_ids_in; 2220827bd09bSSatish Balay msg_ids_out = ids_out = gs->msg_ids_out; 2221827bd09bSSatish Balay dptr2 = gs->out; 2222827bd09bSSatish Balay in1=in2 = gs->in; 2223827bd09bSSatish Balay 2224827bd09bSSatish Balay /* post the receives */ 2225827bd09bSSatish Balay /* msg_nodes=nodes; */ 2226827bd09bSSatish Balay do 2227827bd09bSSatish Balay { 2228827bd09bSSatish Balay /* Should MPI_ANY_SOURCE be replaced by *list ? In that case do the 2229827bd09bSSatish Balay second one *list and do list++ afterwards */ 22303fdc5746SBarry Smith ierr = MPI_Irecv(in1, *size, MPIU_SCALAR, MPI_ANY_SOURCE, MSGTAG1 + *list++, gs->gs_comm, msg_ids_in++);CHKERRQ(ierr); 2231827bd09bSSatish Balay in1 += *size++; 2232827bd09bSSatish Balay } 2233827bd09bSSatish Balay while (*++msg_nodes); 2234827bd09bSSatish Balay msg_nodes=nodes; 2235827bd09bSSatish Balay 2236827bd09bSSatish Balay /* load gs values into in out gs buffers */ 2237827bd09bSSatish Balay while (*iptr >= 0) 2238827bd09bSSatish Balay {*dptr3++ = *(in_vals + *iptr++);} 2239827bd09bSSatish Balay 2240827bd09bSSatish Balay /* load out buffers and post the sends */ 2241827bd09bSSatish Balay while ((iptr = *msg_nodes++)) 2242827bd09bSSatish Balay { 2243827bd09bSSatish Balay dptr3 = dptr2; 2244827bd09bSSatish Balay while (*iptr >= 0) 2245827bd09bSSatish Balay {*dptr2++ = *(dptr1 + *iptr++);} 2246827bd09bSSatish Balay /* CHECK PERSISTENT COMMS MODE FOR ALL THIS STUFF */ 2247827bd09bSSatish Balay /* is msg_ids_out++ correct? */ 22483fdc5746SBarry Smith ierr = MPI_Isend(dptr3, *msg_size++, MPIU_SCALAR, *msg_list++, MSGTAG1+my_id, gs->gs_comm, msg_ids_out++);CHKERRQ(ierr); 2249827bd09bSSatish Balay } 2250827bd09bSSatish Balay 2251827bd09bSSatish Balay /* process the received data */ 2252827bd09bSSatish Balay if (gs->max_left_over) 2253827bd09bSSatish Balay {gs_gop_tree_min(gs,in_vals);} 2254827bd09bSSatish Balay 2255827bd09bSSatish Balay msg_nodes=nodes; 2256827bd09bSSatish Balay while ((iptr = *nodes++)) 2257827bd09bSSatish Balay { 2258827bd09bSSatish Balay /* Should I check the return value of MPI_Wait() or status? */ 2259827bd09bSSatish Balay /* Can this loop be replaced by a call to MPI_Waitall()? */ 22603fdc5746SBarry Smith ierr = MPI_Wait(ids_in++, &status);CHKERRQ(ierr); 2261827bd09bSSatish Balay while (*iptr >= 0) 226239945688SSatish Balay {*(dptr1 + *iptr) = PetscMin(*(dptr1 + *iptr),*in2); iptr++; in2++;} 2263827bd09bSSatish Balay } 2264827bd09bSSatish Balay 2265827bd09bSSatish Balay /* replace vals */ 2266827bd09bSSatish Balay while (*pw >= 0) 2267827bd09bSSatish Balay {*(in_vals + *pw++) = *dptr1++;} 2268827bd09bSSatish Balay 2269827bd09bSSatish Balay /* clear isend message handles */ 2270827bd09bSSatish Balay /* This changed for clarity though it could be the same */ 2271827bd09bSSatish Balay while (*msg_nodes++) 2272827bd09bSSatish Balay /* Should I check the return value of MPI_Wait() or status? */ 2273827bd09bSSatish Balay /* Can this loop be replaced by a call to MPI_Waitall()? */ 22743fdc5746SBarry Smith {ierr = MPI_Wait(ids_out++, &status);CHKERRQ(ierr);} 22753fdc5746SBarry Smith PetscFunctionReturn(0); 2276827bd09bSSatish Balay } 2277827bd09bSSatish Balay 22787b1ae94cSBarry Smith /******************************************************************************/ 22790924e98cSBarry Smith static PetscErrorCode gs_gop_tree_min(gs_id *gs, PetscScalar *vals) 2280827bd09bSSatish Balay { 228152f87cdaSBarry Smith PetscInt size; 228252f87cdaSBarry Smith PetscInt *in, *out; 2283a501084fSBarry Smith PetscScalar *buf, *work; 22843fdc5746SBarry Smith PetscErrorCode ierr; 2285827bd09bSSatish Balay 22863fdc5746SBarry Smith PetscFunctionBegin; 2287827bd09bSSatish Balay in = gs->tree_map_in; 2288827bd09bSSatish Balay out = gs->tree_map_out; 2289827bd09bSSatish Balay buf = gs->tree_buf; 2290827bd09bSSatish Balay work = gs->tree_work; 2291827bd09bSSatish Balay size = gs->tree_nel; 2292827bd09bSSatish Balay 2293827bd09bSSatish Balay rvec_set(buf,REAL_MAX,size); 2294827bd09bSSatish Balay 2295827bd09bSSatish Balay while (*in >= 0) 2296827bd09bSSatish Balay {*(buf + *out++) = *(vals + *in++);} 2297827bd09bSSatish Balay 2298827bd09bSSatish Balay in = gs->tree_map_in; 2299827bd09bSSatish Balay out = gs->tree_map_out; 23003fdc5746SBarry Smith ierr = MPI_Allreduce(buf,work,size,MPIU_SCALAR,MPI_MIN,gs->gs_comm);CHKERRQ(ierr); 2301827bd09bSSatish Balay while (*in >= 0) 2302827bd09bSSatish Balay {*(vals + *in++) = *(work + *out++);} 23033fdc5746SBarry Smith PetscFunctionReturn(0); 2304827bd09bSSatish Balay } 2305827bd09bSSatish Balay 23067b1ae94cSBarry Smith /******************************************************************************/ 23070924e98cSBarry Smith static PetscErrorCode gs_gop_times( gs_id *gs, PetscScalar *vals) 2308827bd09bSSatish Balay { 23093fdc5746SBarry Smith PetscFunctionBegin; 2310827bd09bSSatish Balay /* local only operations!!! */ 2311827bd09bSSatish Balay if (gs->num_local) 2312827bd09bSSatish Balay {gs_gop_local_times(gs,vals);} 2313827bd09bSSatish Balay 2314827bd09bSSatish Balay /* if intersection tree/pairwise and local isn't empty */ 2315827bd09bSSatish Balay if (gs->num_local_gop) 2316827bd09bSSatish Balay { 2317827bd09bSSatish Balay gs_gop_local_in_times(gs,vals); 2318827bd09bSSatish Balay 2319827bd09bSSatish Balay /* pairwise */ 2320827bd09bSSatish Balay if (gs->num_pairs) 2321827bd09bSSatish Balay {gs_gop_pairwise_times(gs,vals);} 2322827bd09bSSatish Balay 2323827bd09bSSatish Balay /* tree */ 2324827bd09bSSatish Balay else if (gs->max_left_over) 2325827bd09bSSatish Balay {gs_gop_tree_times(gs,vals);} 2326827bd09bSSatish Balay 2327827bd09bSSatish Balay gs_gop_local_out(gs,vals); 2328827bd09bSSatish Balay } 2329827bd09bSSatish Balay /* if intersection tree/pairwise and local is empty */ 2330827bd09bSSatish Balay else 2331827bd09bSSatish Balay { 2332827bd09bSSatish Balay /* pairwise */ 2333827bd09bSSatish Balay if (gs->num_pairs) 2334827bd09bSSatish Balay {gs_gop_pairwise_times(gs,vals);} 2335827bd09bSSatish Balay 2336827bd09bSSatish Balay /* tree */ 2337827bd09bSSatish Balay else if (gs->max_left_over) 2338827bd09bSSatish Balay {gs_gop_tree_times(gs,vals);} 2339827bd09bSSatish Balay } 23403fdc5746SBarry Smith PetscFunctionReturn(0); 2341827bd09bSSatish Balay } 2342827bd09bSSatish Balay 23437b1ae94cSBarry Smith /******************************************************************************/ 23440924e98cSBarry Smith static PetscErrorCode gs_gop_local_times( gs_id *gs, PetscScalar *vals) 2345827bd09bSSatish Balay { 234652f87cdaSBarry Smith PetscInt *num, *map, **reduce; 2347a501084fSBarry Smith PetscScalar tmp; 2348827bd09bSSatish Balay 23493fdc5746SBarry Smith PetscFunctionBegin; 2350827bd09bSSatish Balay num = gs->num_local_reduce; 2351827bd09bSSatish Balay reduce = gs->local_reduce; 2352827bd09bSSatish Balay while ((map = *reduce)) 2353827bd09bSSatish Balay { 2354827bd09bSSatish Balay /* wall */ 2355827bd09bSSatish Balay if (*num == 2) 2356827bd09bSSatish Balay { 2357827bd09bSSatish Balay num ++; reduce++; 2358827bd09bSSatish Balay vals[map[1]] = vals[map[0]] *= vals[map[1]]; 2359827bd09bSSatish Balay } 2360827bd09bSSatish Balay /* corner shared by three elements */ 2361827bd09bSSatish Balay else if (*num == 3) 2362827bd09bSSatish Balay { 2363827bd09bSSatish Balay num ++; reduce++; 2364827bd09bSSatish Balay vals[map[2]]=vals[map[1]]=vals[map[0]]*=(vals[map[1]]*vals[map[2]]); 2365827bd09bSSatish Balay } 2366827bd09bSSatish Balay /* corner shared by four elements */ 2367827bd09bSSatish Balay else if (*num == 4) 2368827bd09bSSatish Balay { 2369827bd09bSSatish Balay num ++; reduce++; 2370827bd09bSSatish Balay vals[map[1]]=vals[map[2]]=vals[map[3]]=vals[map[0]] *= 2371827bd09bSSatish Balay (vals[map[1]] * vals[map[2]] * vals[map[3]]); 2372827bd09bSSatish Balay } 2373827bd09bSSatish Balay /* general case ... odd geoms ... 3D*/ 2374827bd09bSSatish Balay else 2375827bd09bSSatish Balay { 2376827bd09bSSatish Balay num ++; 2377827bd09bSSatish Balay tmp = 1.0; 2378827bd09bSSatish Balay while (*map >= 0) 2379827bd09bSSatish Balay {tmp *= *(vals + *map++);} 2380827bd09bSSatish Balay 2381827bd09bSSatish Balay map = *reduce++; 2382827bd09bSSatish Balay while (*map >= 0) 2383827bd09bSSatish Balay {*(vals + *map++) = tmp;} 2384827bd09bSSatish Balay } 2385827bd09bSSatish Balay } 23863fdc5746SBarry Smith PetscFunctionReturn(0); 2387827bd09bSSatish Balay } 2388827bd09bSSatish Balay 23897b1ae94cSBarry Smith /******************************************************************************/ 23900924e98cSBarry Smith static PetscErrorCode gs_gop_local_in_times( gs_id *gs, PetscScalar *vals) 2391827bd09bSSatish Balay { 239252f87cdaSBarry Smith PetscInt *num, *map, **reduce; 2393a501084fSBarry Smith PetscScalar *base; 2394827bd09bSSatish Balay 23953fdc5746SBarry Smith PetscFunctionBegin; 2396827bd09bSSatish Balay num = gs->num_gop_local_reduce; 2397827bd09bSSatish Balay reduce = gs->gop_local_reduce; 2398827bd09bSSatish Balay while ((map = *reduce++)) 2399827bd09bSSatish Balay { 2400827bd09bSSatish Balay /* wall */ 2401827bd09bSSatish Balay if (*num == 2) 2402827bd09bSSatish Balay { 2403827bd09bSSatish Balay num ++; 2404827bd09bSSatish Balay vals[map[0]] *= vals[map[1]]; 2405827bd09bSSatish Balay } 2406827bd09bSSatish Balay /* corner shared by three elements */ 2407827bd09bSSatish Balay else if (*num == 3) 2408827bd09bSSatish Balay { 2409827bd09bSSatish Balay num ++; 2410827bd09bSSatish Balay vals[map[0]] *= (vals[map[1]] * vals[map[2]]); 2411827bd09bSSatish Balay } 2412827bd09bSSatish Balay /* corner shared by four elements */ 2413827bd09bSSatish Balay else if (*num == 4) 2414827bd09bSSatish Balay { 2415827bd09bSSatish Balay num ++; 2416827bd09bSSatish Balay vals[map[0]] *= (vals[map[1]] * vals[map[2]] * vals[map[3]]); 2417827bd09bSSatish Balay } 2418827bd09bSSatish Balay /* general case ... odd geoms ... 3D*/ 2419827bd09bSSatish Balay else 2420827bd09bSSatish Balay { 2421827bd09bSSatish Balay num++; 2422827bd09bSSatish Balay base = vals + *map++; 2423827bd09bSSatish Balay while (*map >= 0) 2424827bd09bSSatish Balay {*base *= *(vals + *map++);} 2425827bd09bSSatish Balay } 2426827bd09bSSatish Balay } 24273fdc5746SBarry Smith PetscFunctionReturn(0); 2428827bd09bSSatish Balay } 2429827bd09bSSatish Balay 24307b1ae94cSBarry Smith /******************************************************************************/ 24310924e98cSBarry Smith static PetscErrorCode gs_gop_pairwise_times( gs_id *gs, PetscScalar *in_vals) 2432827bd09bSSatish Balay { 2433a501084fSBarry Smith PetscScalar *dptr1, *dptr2, *dptr3, *in1, *in2; 243452f87cdaSBarry Smith PetscInt *iptr, *msg_list, *msg_size, **msg_nodes; 243552f87cdaSBarry Smith PetscInt *pw, *list, *size, **nodes; 2436827bd09bSSatish Balay MPI_Request *msg_ids_in, *msg_ids_out, *ids_in, *ids_out; 2437827bd09bSSatish Balay MPI_Status status; 24383fdc5746SBarry Smith PetscErrorCode ierr; 2439827bd09bSSatish Balay 24403fdc5746SBarry Smith PetscFunctionBegin; 2441a501084fSBarry Smith /* strip and load s */ 2442827bd09bSSatish Balay msg_list =list = gs->pair_list; 2443827bd09bSSatish Balay msg_size =size = gs->msg_sizes; 2444827bd09bSSatish Balay msg_nodes=nodes = gs->node_list; 2445827bd09bSSatish Balay iptr=pw = gs->pw_elm_list; 2446827bd09bSSatish Balay dptr1=dptr3 = gs->pw_vals; 2447827bd09bSSatish Balay msg_ids_in = ids_in = gs->msg_ids_in; 2448827bd09bSSatish Balay msg_ids_out = ids_out = gs->msg_ids_out; 2449827bd09bSSatish Balay dptr2 = gs->out; 2450827bd09bSSatish Balay in1=in2 = gs->in; 2451827bd09bSSatish Balay 2452827bd09bSSatish Balay /* post the receives */ 2453827bd09bSSatish Balay /* msg_nodes=nodes; */ 2454827bd09bSSatish Balay do 2455827bd09bSSatish Balay { 2456827bd09bSSatish Balay /* Should MPI_ANY_SOURCE be replaced by *list ? In that case do the 2457827bd09bSSatish Balay second one *list and do list++ afterwards */ 24583fdc5746SBarry Smith ierr = MPI_Irecv(in1, *size, MPIU_SCALAR, MPI_ANY_SOURCE, MSGTAG1 + *list++, gs->gs_comm, msg_ids_in++);CHKERRQ(ierr); 2459827bd09bSSatish Balay in1 += *size++; 2460827bd09bSSatish Balay } 2461827bd09bSSatish Balay while (*++msg_nodes); 2462827bd09bSSatish Balay msg_nodes=nodes; 2463827bd09bSSatish Balay 2464827bd09bSSatish Balay /* load gs values into in out gs buffers */ 2465827bd09bSSatish Balay while (*iptr >= 0) 2466827bd09bSSatish Balay {*dptr3++ = *(in_vals + *iptr++);} 2467827bd09bSSatish Balay 2468827bd09bSSatish Balay /* load out buffers and post the sends */ 2469827bd09bSSatish Balay while ((iptr = *msg_nodes++)) 2470827bd09bSSatish Balay { 2471827bd09bSSatish Balay dptr3 = dptr2; 2472827bd09bSSatish Balay while (*iptr >= 0) 2473827bd09bSSatish Balay {*dptr2++ = *(dptr1 + *iptr++);} 2474827bd09bSSatish Balay /* CHECK PERSISTENT COMMS MODE FOR ALL THIS STUFF */ 2475827bd09bSSatish Balay /* is msg_ids_out++ correct? */ 24763fdc5746SBarry Smith ierr = MPI_Isend(dptr3, *msg_size++, MPIU_SCALAR, *msg_list++, MSGTAG1+my_id, gs->gs_comm, msg_ids_out++);CHKERRQ(ierr); 2477827bd09bSSatish Balay } 2478827bd09bSSatish Balay 2479827bd09bSSatish Balay if (gs->max_left_over) 2480827bd09bSSatish Balay {gs_gop_tree_times(gs,in_vals);} 2481827bd09bSSatish Balay 2482827bd09bSSatish Balay /* process the received data */ 2483827bd09bSSatish Balay msg_nodes=nodes; 2484827bd09bSSatish Balay while ((iptr = *nodes++)) 2485827bd09bSSatish Balay { 2486827bd09bSSatish Balay /* Should I check the return value of MPI_Wait() or status? */ 2487827bd09bSSatish Balay /* Can this loop be replaced by a call to MPI_Waitall()? */ 24883fdc5746SBarry Smith ierr = MPI_Wait(ids_in++, &status);CHKERRQ(ierr); 2489827bd09bSSatish Balay while (*iptr >= 0) 2490827bd09bSSatish Balay {*(dptr1 + *iptr++) *= *in2++;} 2491827bd09bSSatish Balay } 2492827bd09bSSatish Balay 2493827bd09bSSatish Balay /* replace vals */ 2494827bd09bSSatish Balay while (*pw >= 0) 2495827bd09bSSatish Balay {*(in_vals + *pw++) = *dptr1++;} 2496827bd09bSSatish Balay 2497827bd09bSSatish Balay /* clear isend message handles */ 2498827bd09bSSatish Balay /* This changed for clarity though it could be the same */ 2499827bd09bSSatish Balay while (*msg_nodes++) 2500827bd09bSSatish Balay /* Should I check the return value of MPI_Wait() or status? */ 2501827bd09bSSatish Balay /* Can this loop be replaced by a call to MPI_Waitall()? */ 25023fdc5746SBarry Smith {ierr = MPI_Wait(ids_out++, &status);CHKERRQ(ierr);} 25033fdc5746SBarry Smith PetscFunctionReturn(0); 2504827bd09bSSatish Balay } 2505827bd09bSSatish Balay 25067b1ae94cSBarry Smith /******************************************************************************/ 25070924e98cSBarry Smith static PetscErrorCode gs_gop_tree_times(gs_id *gs, PetscScalar *vals) 2508827bd09bSSatish Balay { 250952f87cdaSBarry Smith PetscInt size; 251052f87cdaSBarry Smith PetscInt *in, *out; 2511a501084fSBarry Smith PetscScalar *buf, *work; 25123fdc5746SBarry Smith PetscErrorCode ierr; 2513827bd09bSSatish Balay 25143fdc5746SBarry Smith PetscFunctionBegin; 2515827bd09bSSatish Balay in = gs->tree_map_in; 2516827bd09bSSatish Balay out = gs->tree_map_out; 2517827bd09bSSatish Balay buf = gs->tree_buf; 2518827bd09bSSatish Balay work = gs->tree_work; 2519827bd09bSSatish Balay size = gs->tree_nel; 2520827bd09bSSatish Balay 2521827bd09bSSatish Balay rvec_one(buf,size); 2522827bd09bSSatish Balay 2523827bd09bSSatish Balay while (*in >= 0) 2524827bd09bSSatish Balay {*(buf + *out++) = *(vals + *in++);} 2525827bd09bSSatish Balay 2526827bd09bSSatish Balay in = gs->tree_map_in; 2527827bd09bSSatish Balay out = gs->tree_map_out; 25283fdc5746SBarry Smith ierr = MPI_Allreduce(buf,work,size,MPIU_SCALAR,MPI_PROD,gs->gs_comm);CHKERRQ(ierr); 2529827bd09bSSatish Balay while (*in >= 0) 2530827bd09bSSatish Balay {*(vals + *in++) = *(work + *out++);} 25313fdc5746SBarry Smith PetscFunctionReturn(0); 2532827bd09bSSatish Balay } 2533827bd09bSSatish Balay 25347b1ae94cSBarry Smith /******************************************************************************/ 25350924e98cSBarry Smith static PetscErrorCode gs_gop_plus( gs_id *gs, PetscScalar *vals) 2536827bd09bSSatish Balay { 25373fdc5746SBarry Smith PetscFunctionBegin; 2538827bd09bSSatish Balay /* local only operations!!! */ 2539827bd09bSSatish Balay if (gs->num_local) 2540827bd09bSSatish Balay {gs_gop_local_plus(gs,vals);} 2541827bd09bSSatish Balay 2542827bd09bSSatish Balay /* if intersection tree/pairwise and local isn't empty */ 2543827bd09bSSatish Balay if (gs->num_local_gop) 2544827bd09bSSatish Balay { 2545827bd09bSSatish Balay gs_gop_local_in_plus(gs,vals); 2546827bd09bSSatish Balay 2547827bd09bSSatish Balay /* pairwise will NOT do tree inside ... */ 2548827bd09bSSatish Balay if (gs->num_pairs) 2549827bd09bSSatish Balay {gs_gop_pairwise_plus(gs,vals);} 2550827bd09bSSatish Balay 2551827bd09bSSatish Balay /* tree */ 2552827bd09bSSatish Balay if (gs->max_left_over) 2553827bd09bSSatish Balay {gs_gop_tree_plus(gs,vals);} 2554827bd09bSSatish Balay 2555827bd09bSSatish Balay gs_gop_local_out(gs,vals); 2556827bd09bSSatish Balay } 2557827bd09bSSatish Balay /* if intersection tree/pairwise and local is empty */ 2558827bd09bSSatish Balay else 2559827bd09bSSatish Balay { 2560827bd09bSSatish Balay /* pairwise will NOT do tree inside */ 2561827bd09bSSatish Balay if (gs->num_pairs) 2562827bd09bSSatish Balay {gs_gop_pairwise_plus(gs,vals);} 2563827bd09bSSatish Balay 2564827bd09bSSatish Balay /* tree */ 2565827bd09bSSatish Balay if (gs->max_left_over) 2566827bd09bSSatish Balay {gs_gop_tree_plus(gs,vals);} 2567827bd09bSSatish Balay } 25683fdc5746SBarry Smith PetscFunctionReturn(0); 2569827bd09bSSatish Balay } 2570827bd09bSSatish Balay 25717b1ae94cSBarry Smith /******************************************************************************/ 25720924e98cSBarry Smith static PetscErrorCode gs_gop_local_plus( gs_id *gs, PetscScalar *vals) 2573827bd09bSSatish Balay { 257452f87cdaSBarry Smith PetscInt *num, *map, **reduce; 2575a501084fSBarry Smith PetscScalar tmp; 2576827bd09bSSatish Balay 25773fdc5746SBarry Smith PetscFunctionBegin; 2578827bd09bSSatish Balay num = gs->num_local_reduce; 2579827bd09bSSatish Balay reduce = gs->local_reduce; 2580827bd09bSSatish Balay while ((map = *reduce)) 2581827bd09bSSatish Balay { 2582827bd09bSSatish Balay /* wall */ 2583827bd09bSSatish Balay if (*num == 2) 2584827bd09bSSatish Balay { 2585827bd09bSSatish Balay num ++; reduce++; 2586827bd09bSSatish Balay vals[map[1]] = vals[map[0]] += vals[map[1]]; 2587827bd09bSSatish Balay } 2588827bd09bSSatish Balay /* corner shared by three elements */ 2589827bd09bSSatish Balay else if (*num == 3) 2590827bd09bSSatish Balay { 2591827bd09bSSatish Balay num ++; reduce++; 2592827bd09bSSatish Balay vals[map[2]]=vals[map[1]]=vals[map[0]]+=(vals[map[1]]+vals[map[2]]); 2593827bd09bSSatish Balay } 2594827bd09bSSatish Balay /* corner shared by four elements */ 2595827bd09bSSatish Balay else if (*num == 4) 2596827bd09bSSatish Balay { 2597827bd09bSSatish Balay num ++; reduce++; 2598827bd09bSSatish Balay vals[map[1]]=vals[map[2]]=vals[map[3]]=vals[map[0]] += 2599827bd09bSSatish Balay (vals[map[1]] + vals[map[2]] + vals[map[3]]); 2600827bd09bSSatish Balay } 2601827bd09bSSatish Balay /* general case ... odd geoms ... 3D*/ 2602827bd09bSSatish Balay else 2603827bd09bSSatish Balay { 2604827bd09bSSatish Balay num ++; 2605827bd09bSSatish Balay tmp = 0.0; 2606827bd09bSSatish Balay while (*map >= 0) 2607827bd09bSSatish Balay {tmp += *(vals + *map++);} 2608827bd09bSSatish Balay 2609827bd09bSSatish Balay map = *reduce++; 2610827bd09bSSatish Balay while (*map >= 0) 2611827bd09bSSatish Balay {*(vals + *map++) = tmp;} 2612827bd09bSSatish Balay } 2613827bd09bSSatish Balay } 26143fdc5746SBarry Smith PetscFunctionReturn(0); 2615827bd09bSSatish Balay } 2616827bd09bSSatish Balay 26177b1ae94cSBarry Smith /******************************************************************************/ 26180924e98cSBarry Smith static PetscErrorCode gs_gop_local_in_plus( gs_id *gs, PetscScalar *vals) 2619827bd09bSSatish Balay { 262052f87cdaSBarry Smith PetscInt *num, *map, **reduce; 2621a501084fSBarry Smith PetscScalar *base; 2622827bd09bSSatish Balay 26233fdc5746SBarry Smith PetscFunctionBegin; 2624827bd09bSSatish Balay num = gs->num_gop_local_reduce; 2625827bd09bSSatish Balay reduce = gs->gop_local_reduce; 2626827bd09bSSatish Balay while ((map = *reduce++)) 2627827bd09bSSatish Balay { 2628827bd09bSSatish Balay /* wall */ 2629827bd09bSSatish Balay if (*num == 2) 2630827bd09bSSatish Balay { 2631827bd09bSSatish Balay num ++; 2632827bd09bSSatish Balay vals[map[0]] += vals[map[1]]; 2633827bd09bSSatish Balay } 2634827bd09bSSatish Balay /* corner shared by three elements */ 2635827bd09bSSatish Balay else if (*num == 3) 2636827bd09bSSatish Balay { 2637827bd09bSSatish Balay num ++; 2638827bd09bSSatish Balay vals[map[0]] += (vals[map[1]] + vals[map[2]]); 2639827bd09bSSatish Balay } 2640827bd09bSSatish Balay /* corner shared by four elements */ 2641827bd09bSSatish Balay else if (*num == 4) 2642827bd09bSSatish Balay { 2643827bd09bSSatish Balay num ++; 2644827bd09bSSatish Balay vals[map[0]] += (vals[map[1]] + vals[map[2]] + vals[map[3]]); 2645827bd09bSSatish Balay } 2646827bd09bSSatish Balay /* general case ... odd geoms ... 3D*/ 2647827bd09bSSatish Balay else 2648827bd09bSSatish Balay { 2649827bd09bSSatish Balay num++; 2650827bd09bSSatish Balay base = vals + *map++; 2651827bd09bSSatish Balay while (*map >= 0) 2652827bd09bSSatish Balay {*base += *(vals + *map++);} 2653827bd09bSSatish Balay } 2654827bd09bSSatish Balay } 26553fdc5746SBarry Smith PetscFunctionReturn(0); 2656827bd09bSSatish Balay } 2657827bd09bSSatish Balay 26587b1ae94cSBarry Smith /******************************************************************************/ 26590924e98cSBarry Smith static PetscErrorCode gs_gop_pairwise_plus( gs_id *gs, PetscScalar *in_vals) 2660827bd09bSSatish Balay { 2661a501084fSBarry Smith PetscScalar *dptr1, *dptr2, *dptr3, *in1, *in2; 266252f87cdaSBarry Smith PetscInt *iptr, *msg_list, *msg_size, **msg_nodes; 266352f87cdaSBarry Smith PetscInt *pw, *list, *size, **nodes; 2664827bd09bSSatish Balay MPI_Request *msg_ids_in, *msg_ids_out, *ids_in, *ids_out; 2665827bd09bSSatish Balay MPI_Status status; 26663fdc5746SBarry Smith PetscErrorCode ierr; 2667827bd09bSSatish Balay 26683fdc5746SBarry Smith PetscFunctionBegin; 2669a501084fSBarry Smith /* strip and load s */ 2670827bd09bSSatish Balay msg_list =list = gs->pair_list; 2671827bd09bSSatish Balay msg_size =size = gs->msg_sizes; 2672827bd09bSSatish Balay msg_nodes=nodes = gs->node_list; 2673827bd09bSSatish Balay iptr=pw = gs->pw_elm_list; 2674827bd09bSSatish Balay dptr1=dptr3 = gs->pw_vals; 2675827bd09bSSatish Balay msg_ids_in = ids_in = gs->msg_ids_in; 2676827bd09bSSatish Balay msg_ids_out = ids_out = gs->msg_ids_out; 2677827bd09bSSatish Balay dptr2 = gs->out; 2678827bd09bSSatish Balay in1=in2 = gs->in; 2679827bd09bSSatish Balay 2680827bd09bSSatish Balay /* post the receives */ 2681827bd09bSSatish Balay /* msg_nodes=nodes; */ 2682827bd09bSSatish Balay do 2683827bd09bSSatish Balay { 2684827bd09bSSatish Balay /* Should MPI_ANY_SOURCE be replaced by *list ? In that case do the 2685827bd09bSSatish Balay second one *list and do list++ afterwards */ 26863fdc5746SBarry Smith ierr = MPI_Irecv(in1, *size, MPIU_SCALAR, MPI_ANY_SOURCE, MSGTAG1 + *list++, gs->gs_comm, msg_ids_in++);CHKERRQ(ierr); 2687827bd09bSSatish Balay in1 += *size++; 2688827bd09bSSatish Balay } 2689827bd09bSSatish Balay while (*++msg_nodes); 2690827bd09bSSatish Balay msg_nodes=nodes; 2691827bd09bSSatish Balay 2692827bd09bSSatish Balay /* load gs values into in out gs buffers */ 2693827bd09bSSatish Balay while (*iptr >= 0) 2694827bd09bSSatish Balay {*dptr3++ = *(in_vals + *iptr++);} 2695827bd09bSSatish Balay 2696827bd09bSSatish Balay /* load out buffers and post the sends */ 2697827bd09bSSatish Balay while ((iptr = *msg_nodes++)) 2698827bd09bSSatish Balay { 2699827bd09bSSatish Balay dptr3 = dptr2; 2700827bd09bSSatish Balay while (*iptr >= 0) 2701827bd09bSSatish Balay {*dptr2++ = *(dptr1 + *iptr++);} 2702827bd09bSSatish Balay /* CHECK PERSISTENT COMMS MODE FOR ALL THIS STUFF */ 2703827bd09bSSatish Balay /* is msg_ids_out++ correct? */ 27043fdc5746SBarry Smith ierr = MPI_Isend(dptr3, *msg_size++, MPIU_SCALAR, *msg_list++, MSGTAG1+my_id, gs->gs_comm, msg_ids_out++);CHKERRQ(ierr); 2705827bd09bSSatish Balay } 2706827bd09bSSatish Balay 2707827bd09bSSatish Balay /* do the tree while we're waiting */ 2708827bd09bSSatish Balay if (gs->max_left_over) 2709827bd09bSSatish Balay {gs_gop_tree_plus(gs,in_vals);} 2710827bd09bSSatish Balay 2711827bd09bSSatish Balay /* process the received data */ 2712827bd09bSSatish Balay msg_nodes=nodes; 2713827bd09bSSatish Balay while ((iptr = *nodes++)) 2714827bd09bSSatish Balay { 2715827bd09bSSatish Balay /* Should I check the return value of MPI_Wait() or status? */ 2716827bd09bSSatish Balay /* Can this loop be replaced by a call to MPI_Waitall()? */ 27173fdc5746SBarry Smith ierr = MPI_Wait(ids_in++, &status);CHKERRQ(ierr); 2718827bd09bSSatish Balay while (*iptr >= 0) 2719827bd09bSSatish Balay {*(dptr1 + *iptr++) += *in2++;} 2720827bd09bSSatish Balay } 2721827bd09bSSatish Balay 2722827bd09bSSatish Balay /* replace vals */ 2723827bd09bSSatish Balay while (*pw >= 0) 2724827bd09bSSatish Balay {*(in_vals + *pw++) = *dptr1++;} 2725827bd09bSSatish Balay 2726827bd09bSSatish Balay /* clear isend message handles */ 2727827bd09bSSatish Balay /* This changed for clarity though it could be the same */ 2728827bd09bSSatish Balay while (*msg_nodes++) 2729827bd09bSSatish Balay /* Should I check the return value of MPI_Wait() or status? */ 2730827bd09bSSatish Balay /* Can this loop be replaced by a call to MPI_Waitall()? */ 27313fdc5746SBarry Smith {ierr = MPI_Wait(ids_out++, &status);CHKERRQ(ierr);} 27323fdc5746SBarry Smith PetscFunctionReturn(0); 2733827bd09bSSatish Balay } 2734827bd09bSSatish Balay 27357b1ae94cSBarry Smith /******************************************************************************/ 27360924e98cSBarry Smith static PetscErrorCode gs_gop_tree_plus(gs_id *gs, PetscScalar *vals) 2737827bd09bSSatish Balay { 273852f87cdaSBarry Smith PetscInt size; 273952f87cdaSBarry Smith PetscInt *in, *out; 2740a501084fSBarry Smith PetscScalar *buf, *work; 27413fdc5746SBarry Smith PetscErrorCode ierr; 2742827bd09bSSatish Balay 27433fdc5746SBarry Smith PetscFunctionBegin; 2744827bd09bSSatish Balay in = gs->tree_map_in; 2745827bd09bSSatish Balay out = gs->tree_map_out; 2746827bd09bSSatish Balay buf = gs->tree_buf; 2747827bd09bSSatish Balay work = gs->tree_work; 2748827bd09bSSatish Balay size = gs->tree_nel; 2749827bd09bSSatish Balay 2750827bd09bSSatish Balay rvec_zero(buf,size); 2751827bd09bSSatish Balay 2752827bd09bSSatish Balay while (*in >= 0) 2753827bd09bSSatish Balay {*(buf + *out++) = *(vals + *in++);} 2754827bd09bSSatish Balay 2755827bd09bSSatish Balay in = gs->tree_map_in; 2756827bd09bSSatish Balay out = gs->tree_map_out; 27573fdc5746SBarry Smith ierr = MPI_Allreduce(buf,work,size,MPIU_SCALAR,MPI_SUM,gs->gs_comm);CHKERRQ(ierr); 2758827bd09bSSatish Balay while (*in >= 0) 2759827bd09bSSatish Balay {*(vals + *in++) = *(work + *out++);} 27603fdc5746SBarry Smith PetscFunctionReturn(0); 2761827bd09bSSatish Balay } 2762827bd09bSSatish Balay 27637b1ae94cSBarry Smith /******************************************************************************/ 27640924e98cSBarry Smith PetscErrorCode gs_free( gs_id *gs) 2765827bd09bSSatish Balay { 276652f87cdaSBarry Smith PetscInt i; 2767827bd09bSSatish Balay 27683fdc5746SBarry Smith PetscFunctionBegin; 2769a501084fSBarry Smith if (gs->nghs) {free((void*) gs->nghs);} 2770a501084fSBarry Smith if (gs->pw_nghs) {free((void*) gs->pw_nghs);} 2771827bd09bSSatish Balay 2772827bd09bSSatish Balay /* tree */ 2773827bd09bSSatish Balay if (gs->max_left_over) 2774827bd09bSSatish Balay { 2775a501084fSBarry Smith if (gs->tree_elms) {free((void*) gs->tree_elms);} 2776a501084fSBarry Smith if (gs->tree_buf) {free((void*) gs->tree_buf);} 2777a501084fSBarry Smith if (gs->tree_work) {free((void*) gs->tree_work);} 2778a501084fSBarry Smith if (gs->tree_map_in) {free((void*) gs->tree_map_in);} 2779a501084fSBarry Smith if (gs->tree_map_out) {free((void*) gs->tree_map_out);} 2780827bd09bSSatish Balay } 2781827bd09bSSatish Balay 2782827bd09bSSatish Balay /* pairwise info */ 2783827bd09bSSatish Balay if (gs->num_pairs) 2784827bd09bSSatish Balay { 2785827bd09bSSatish Balay /* should be NULL already */ 2786a501084fSBarry Smith if (gs->ngh_buf) {free((void*) gs->ngh_buf);} 2787a501084fSBarry Smith if (gs->elms) {free((void*) gs->elms);} 2788a501084fSBarry Smith if (gs->local_elms) {free((void*) gs->local_elms);} 2789a501084fSBarry Smith if (gs->companion) {free((void*) gs->companion);} 2790827bd09bSSatish Balay 2791827bd09bSSatish Balay /* only set if pairwise */ 2792a501084fSBarry Smith if (gs->vals) {free((void*) gs->vals);} 2793a501084fSBarry Smith if (gs->in) {free((void*) gs->in);} 2794a501084fSBarry Smith if (gs->out) {free((void*) gs->out);} 2795a501084fSBarry Smith if (gs->msg_ids_in) {free((void*) gs->msg_ids_in);} 2796a501084fSBarry Smith if (gs->msg_ids_out) {free((void*) gs->msg_ids_out);} 2797a501084fSBarry Smith if (gs->pw_vals) {free((void*) gs->pw_vals);} 2798a501084fSBarry Smith if (gs->pw_elm_list) {free((void*) gs->pw_elm_list);} 2799827bd09bSSatish Balay if (gs->node_list) 2800827bd09bSSatish Balay { 2801827bd09bSSatish Balay for (i=0;i<gs->num_pairs;i++) 2802a501084fSBarry Smith {if (gs->node_list[i]) {free((void*) gs->node_list[i]);}} 2803a501084fSBarry Smith free((void*) gs->node_list); 2804827bd09bSSatish Balay } 2805a501084fSBarry Smith if (gs->msg_sizes) {free((void*) gs->msg_sizes);} 2806a501084fSBarry Smith if (gs->pair_list) {free((void*) gs->pair_list);} 2807827bd09bSSatish Balay } 2808827bd09bSSatish Balay 2809827bd09bSSatish Balay /* local info */ 2810827bd09bSSatish Balay if (gs->num_local_total>=0) 2811827bd09bSSatish Balay { 2812827bd09bSSatish Balay for (i=0;i<gs->num_local_total+1;i++) 2813827bd09bSSatish Balay /* for (i=0;i<gs->num_local_total;i++) */ 2814827bd09bSSatish Balay { 2815827bd09bSSatish Balay if (gs->num_gop_local_reduce[i]) 2816a501084fSBarry Smith {free((void*) gs->gop_local_reduce[i]);} 2817827bd09bSSatish Balay } 2818827bd09bSSatish Balay } 2819827bd09bSSatish Balay 2820827bd09bSSatish Balay /* if intersection tree/pairwise and local isn't empty */ 2821a501084fSBarry Smith if (gs->gop_local_reduce) {free((void*) gs->gop_local_reduce);} 2822a501084fSBarry Smith if (gs->num_gop_local_reduce) {free((void*) gs->num_gop_local_reduce);} 2823827bd09bSSatish Balay 2824a501084fSBarry Smith free((void*) gs); 28253fdc5746SBarry Smith PetscFunctionReturn(0); 2826827bd09bSSatish Balay } 2827827bd09bSSatish Balay 28287b1ae94cSBarry Smith /******************************************************************************/ 282952f87cdaSBarry Smith PetscErrorCode gs_gop_vec( gs_id *gs, PetscScalar *vals, const char *op, PetscInt step) 2830827bd09bSSatish Balay { 2831*d1528f56SBarry Smith PetscErrorCode ierr; 2832*d1528f56SBarry Smith 28333fdc5746SBarry Smith PetscFunctionBegin; 2834827bd09bSSatish Balay switch (*op) { 2835827bd09bSSatish Balay case '+': 2836827bd09bSSatish Balay gs_gop_vec_plus(gs,vals,step); 2837827bd09bSSatish Balay break; 2838827bd09bSSatish Balay default: 2839*d1528f56SBarry Smith ierr = PetscInfo1(0,"gs_gop_vec() :: %c is not a valid op",op[0]); 2840*d1528f56SBarry Smith ierr = PetscInfo(0,"gs_gop_vec() :: default :: plus"); 2841827bd09bSSatish Balay gs_gop_vec_plus(gs,vals,step); 2842827bd09bSSatish Balay break; 2843827bd09bSSatish Balay } 28443fdc5746SBarry Smith PetscFunctionReturn(0); 2845827bd09bSSatish Balay } 2846827bd09bSSatish Balay 28477b1ae94cSBarry Smith /******************************************************************************/ 284852f87cdaSBarry Smith static PetscErrorCode gs_gop_vec_plus( gs_id *gs, PetscScalar *vals, PetscInt step) 2849827bd09bSSatish Balay { 28503fdc5746SBarry Smith PetscFunctionBegin; 2851388eb383SBarry Smith if (!gs) {SETERRQ(PETSC_ERR_PLIB,"gs_gop_vec() passed NULL gs handle!!!");} 2852827bd09bSSatish Balay 2853827bd09bSSatish Balay /* local only operations!!! */ 2854827bd09bSSatish Balay if (gs->num_local) 2855827bd09bSSatish Balay {gs_gop_vec_local_plus(gs,vals,step);} 2856827bd09bSSatish Balay 2857827bd09bSSatish Balay /* if intersection tree/pairwise and local isn't empty */ 2858827bd09bSSatish Balay if (gs->num_local_gop) 2859827bd09bSSatish Balay { 2860827bd09bSSatish Balay gs_gop_vec_local_in_plus(gs,vals,step); 2861827bd09bSSatish Balay 2862827bd09bSSatish Balay /* pairwise */ 2863827bd09bSSatish Balay if (gs->num_pairs) 2864827bd09bSSatish Balay {gs_gop_vec_pairwise_plus(gs,vals,step);} 2865827bd09bSSatish Balay 2866827bd09bSSatish Balay /* tree */ 2867827bd09bSSatish Balay else if (gs->max_left_over) 2868827bd09bSSatish Balay {gs_gop_vec_tree_plus(gs,vals,step);} 2869827bd09bSSatish Balay 2870827bd09bSSatish Balay gs_gop_vec_local_out(gs,vals,step); 2871827bd09bSSatish Balay } 2872827bd09bSSatish Balay /* if intersection tree/pairwise and local is empty */ 2873827bd09bSSatish Balay else 2874827bd09bSSatish Balay { 2875827bd09bSSatish Balay /* pairwise */ 2876827bd09bSSatish Balay if (gs->num_pairs) 2877827bd09bSSatish Balay {gs_gop_vec_pairwise_plus(gs,vals,step);} 2878827bd09bSSatish Balay 2879827bd09bSSatish Balay /* tree */ 2880827bd09bSSatish Balay else if (gs->max_left_over) 2881827bd09bSSatish Balay {gs_gop_vec_tree_plus(gs,vals,step);} 2882827bd09bSSatish Balay } 28833fdc5746SBarry Smith PetscFunctionReturn(0); 2884827bd09bSSatish Balay } 2885827bd09bSSatish Balay 28867b1ae94cSBarry Smith /******************************************************************************/ 288752f87cdaSBarry Smith static PetscErrorCode gs_gop_vec_local_plus( gs_id *gs, PetscScalar *vals, PetscInt step) 2888827bd09bSSatish Balay { 288952f87cdaSBarry Smith PetscInt *num, *map, **reduce; 2890a501084fSBarry Smith PetscScalar *base; 2891827bd09bSSatish Balay 28923fdc5746SBarry Smith PetscFunctionBegin; 2893827bd09bSSatish Balay num = gs->num_local_reduce; 2894827bd09bSSatish Balay reduce = gs->local_reduce; 2895827bd09bSSatish Balay while ((map = *reduce)) 2896827bd09bSSatish Balay { 2897827bd09bSSatish Balay base = vals + map[0] * step; 2898827bd09bSSatish Balay 2899827bd09bSSatish Balay /* wall */ 2900827bd09bSSatish Balay if (*num == 2) 2901827bd09bSSatish Balay { 2902827bd09bSSatish Balay num++; reduce++; 2903827bd09bSSatish Balay rvec_add (base,vals+map[1]*step,step); 2904827bd09bSSatish Balay rvec_copy(vals+map[1]*step,base,step); 2905827bd09bSSatish Balay } 2906827bd09bSSatish Balay /* corner shared by three elements */ 2907827bd09bSSatish Balay else if (*num == 3) 2908827bd09bSSatish Balay { 2909827bd09bSSatish Balay num++; reduce++; 2910827bd09bSSatish Balay rvec_add (base,vals+map[1]*step,step); 2911827bd09bSSatish Balay rvec_add (base,vals+map[2]*step,step); 2912827bd09bSSatish Balay rvec_copy(vals+map[2]*step,base,step); 2913827bd09bSSatish Balay rvec_copy(vals+map[1]*step,base,step); 2914827bd09bSSatish Balay } 2915827bd09bSSatish Balay /* corner shared by four elements */ 2916827bd09bSSatish Balay else if (*num == 4) 2917827bd09bSSatish Balay { 2918827bd09bSSatish Balay num++; reduce++; 2919827bd09bSSatish Balay rvec_add (base,vals+map[1]*step,step); 2920827bd09bSSatish Balay rvec_add (base,vals+map[2]*step,step); 2921827bd09bSSatish Balay rvec_add (base,vals+map[3]*step,step); 2922827bd09bSSatish Balay rvec_copy(vals+map[3]*step,base,step); 2923827bd09bSSatish Balay rvec_copy(vals+map[2]*step,base,step); 2924827bd09bSSatish Balay rvec_copy(vals+map[1]*step,base,step); 2925827bd09bSSatish Balay } 2926827bd09bSSatish Balay /* general case ... odd geoms ... 3D */ 2927827bd09bSSatish Balay else 2928827bd09bSSatish Balay { 2929827bd09bSSatish Balay num++; 2930827bd09bSSatish Balay while (*++map >= 0) 2931827bd09bSSatish Balay {rvec_add (base,vals+*map*step,step);} 2932827bd09bSSatish Balay 2933827bd09bSSatish Balay map = *reduce; 2934827bd09bSSatish Balay while (*++map >= 0) 2935827bd09bSSatish Balay {rvec_copy(vals+*map*step,base,step);} 2936827bd09bSSatish Balay 2937827bd09bSSatish Balay reduce++; 2938827bd09bSSatish Balay } 2939827bd09bSSatish Balay } 29403fdc5746SBarry Smith PetscFunctionReturn(0); 2941827bd09bSSatish Balay } 2942827bd09bSSatish Balay 29437b1ae94cSBarry Smith /******************************************************************************/ 294452f87cdaSBarry Smith static PetscErrorCode gs_gop_vec_local_in_plus( gs_id *gs, PetscScalar *vals, PetscInt step) 2945827bd09bSSatish Balay { 294652f87cdaSBarry Smith PetscInt *num, *map, **reduce; 2947a501084fSBarry Smith PetscScalar *base; 29483fdc5746SBarry Smith PetscFunctionBegin; 2949827bd09bSSatish Balay num = gs->num_gop_local_reduce; 2950827bd09bSSatish Balay reduce = gs->gop_local_reduce; 2951827bd09bSSatish Balay while ((map = *reduce++)) 2952827bd09bSSatish Balay { 2953827bd09bSSatish Balay base = vals + map[0] * step; 2954827bd09bSSatish Balay 2955827bd09bSSatish Balay /* wall */ 2956827bd09bSSatish Balay if (*num == 2) 2957827bd09bSSatish Balay { 2958827bd09bSSatish Balay num ++; 2959827bd09bSSatish Balay rvec_add(base,vals+map[1]*step,step); 2960827bd09bSSatish Balay } 2961827bd09bSSatish Balay /* corner shared by three elements */ 2962827bd09bSSatish Balay else if (*num == 3) 2963827bd09bSSatish Balay { 2964827bd09bSSatish Balay num ++; 2965827bd09bSSatish Balay rvec_add(base,vals+map[1]*step,step); 2966827bd09bSSatish Balay rvec_add(base,vals+map[2]*step,step); 2967827bd09bSSatish Balay } 2968827bd09bSSatish Balay /* corner shared by four elements */ 2969827bd09bSSatish Balay else if (*num == 4) 2970827bd09bSSatish Balay { 2971827bd09bSSatish Balay num ++; 2972827bd09bSSatish Balay rvec_add(base,vals+map[1]*step,step); 2973827bd09bSSatish Balay rvec_add(base,vals+map[2]*step,step); 2974827bd09bSSatish Balay rvec_add(base,vals+map[3]*step,step); 2975827bd09bSSatish Balay } 2976827bd09bSSatish Balay /* general case ... odd geoms ... 3D*/ 2977827bd09bSSatish Balay else 2978827bd09bSSatish Balay { 2979827bd09bSSatish Balay num++; 2980827bd09bSSatish Balay while (*++map >= 0) 2981827bd09bSSatish Balay {rvec_add(base,vals+*map*step,step);} 2982827bd09bSSatish Balay } 2983827bd09bSSatish Balay } 29843fdc5746SBarry Smith PetscFunctionReturn(0); 2985827bd09bSSatish Balay } 2986827bd09bSSatish Balay 29877b1ae94cSBarry Smith /******************************************************************************/ 298852f87cdaSBarry Smith static PetscErrorCode gs_gop_vec_local_out( gs_id *gs, PetscScalar *vals, PetscInt step) 2989827bd09bSSatish Balay { 299052f87cdaSBarry Smith PetscInt *num, *map, **reduce; 2991a501084fSBarry Smith PetscScalar *base; 2992827bd09bSSatish Balay 29933fdc5746SBarry Smith PetscFunctionBegin; 2994827bd09bSSatish Balay num = gs->num_gop_local_reduce; 2995827bd09bSSatish Balay reduce = gs->gop_local_reduce; 2996827bd09bSSatish Balay while ((map = *reduce++)) 2997827bd09bSSatish Balay { 2998827bd09bSSatish Balay base = vals + map[0] * step; 2999827bd09bSSatish Balay 3000827bd09bSSatish Balay /* wall */ 3001827bd09bSSatish Balay if (*num == 2) 3002827bd09bSSatish Balay { 3003827bd09bSSatish Balay num ++; 3004827bd09bSSatish Balay rvec_copy(vals+map[1]*step,base,step); 3005827bd09bSSatish Balay } 3006827bd09bSSatish Balay /* corner shared by three elements */ 3007827bd09bSSatish Balay else if (*num == 3) 3008827bd09bSSatish Balay { 3009827bd09bSSatish Balay num ++; 3010827bd09bSSatish Balay rvec_copy(vals+map[1]*step,base,step); 3011827bd09bSSatish Balay rvec_copy(vals+map[2]*step,base,step); 3012827bd09bSSatish Balay } 3013827bd09bSSatish Balay /* corner shared by four elements */ 3014827bd09bSSatish Balay else if (*num == 4) 3015827bd09bSSatish Balay { 3016827bd09bSSatish Balay num ++; 3017827bd09bSSatish Balay rvec_copy(vals+map[1]*step,base,step); 3018827bd09bSSatish Balay rvec_copy(vals+map[2]*step,base,step); 3019827bd09bSSatish Balay rvec_copy(vals+map[3]*step,base,step); 3020827bd09bSSatish Balay } 3021827bd09bSSatish Balay /* general case ... odd geoms ... 3D*/ 3022827bd09bSSatish Balay else 3023827bd09bSSatish Balay { 3024827bd09bSSatish Balay num++; 3025827bd09bSSatish Balay while (*++map >= 0) 3026827bd09bSSatish Balay {rvec_copy(vals+*map*step,base,step);} 3027827bd09bSSatish Balay } 3028827bd09bSSatish Balay } 30293fdc5746SBarry Smith PetscFunctionReturn(0); 3030827bd09bSSatish Balay } 3031827bd09bSSatish Balay 30327b1ae94cSBarry Smith /******************************************************************************/ 303352f87cdaSBarry Smith static PetscErrorCode gs_gop_vec_pairwise_plus( gs_id *gs, PetscScalar *in_vals, PetscInt step) 3034827bd09bSSatish Balay { 3035a501084fSBarry Smith PetscScalar *dptr1, *dptr2, *dptr3, *in1, *in2; 303652f87cdaSBarry Smith PetscInt *iptr, *msg_list, *msg_size, **msg_nodes; 303752f87cdaSBarry Smith PetscInt *pw, *list, *size, **nodes; 3038827bd09bSSatish Balay MPI_Request *msg_ids_in, *msg_ids_out, *ids_in, *ids_out; 3039827bd09bSSatish Balay MPI_Status status; 3040a501084fSBarry Smith PetscBLASInt i1; 30413fdc5746SBarry Smith PetscErrorCode ierr; 3042827bd09bSSatish Balay 30433fdc5746SBarry Smith PetscFunctionBegin; 3044a501084fSBarry Smith /* strip and load s */ 3045827bd09bSSatish Balay msg_list =list = gs->pair_list; 3046827bd09bSSatish Balay msg_size =size = gs->msg_sizes; 3047827bd09bSSatish Balay msg_nodes=nodes = gs->node_list; 3048827bd09bSSatish Balay iptr=pw = gs->pw_elm_list; 3049827bd09bSSatish Balay dptr1=dptr3 = gs->pw_vals; 3050827bd09bSSatish Balay msg_ids_in = ids_in = gs->msg_ids_in; 3051827bd09bSSatish Balay msg_ids_out = ids_out = gs->msg_ids_out; 3052827bd09bSSatish Balay dptr2 = gs->out; 3053827bd09bSSatish Balay in1=in2 = gs->in; 3054827bd09bSSatish Balay 3055827bd09bSSatish Balay /* post the receives */ 3056827bd09bSSatish Balay /* msg_nodes=nodes; */ 3057827bd09bSSatish Balay do 3058827bd09bSSatish Balay { 3059827bd09bSSatish Balay /* Should MPI_ANY_SOURCE be replaced by *list ? In that case do the 3060827bd09bSSatish Balay second one *list and do list++ afterwards */ 30613fdc5746SBarry Smith ierr = MPI_Irecv(in1, *size *step, MPIU_SCALAR, MPI_ANY_SOURCE, MSGTAG1 + *list++, gs->gs_comm, msg_ids_in++);CHKERRQ(ierr); 3062827bd09bSSatish Balay in1 += *size++ *step; 3063827bd09bSSatish Balay } 3064827bd09bSSatish Balay while (*++msg_nodes); 3065827bd09bSSatish Balay msg_nodes=nodes; 3066827bd09bSSatish Balay 3067827bd09bSSatish Balay /* load gs values into in out gs buffers */ 3068827bd09bSSatish Balay while (*iptr >= 0) 3069827bd09bSSatish Balay { 3070827bd09bSSatish Balay rvec_copy(dptr3,in_vals + *iptr*step,step); 3071827bd09bSSatish Balay dptr3+=step; 3072827bd09bSSatish Balay iptr++; 3073827bd09bSSatish Balay } 3074827bd09bSSatish Balay 3075827bd09bSSatish Balay /* load out buffers and post the sends */ 3076827bd09bSSatish Balay while ((iptr = *msg_nodes++)) 3077827bd09bSSatish Balay { 3078827bd09bSSatish Balay dptr3 = dptr2; 3079827bd09bSSatish Balay while (*iptr >= 0) 3080827bd09bSSatish Balay { 3081827bd09bSSatish Balay rvec_copy(dptr2,dptr1 + *iptr*step,step); 3082827bd09bSSatish Balay dptr2+=step; 3083827bd09bSSatish Balay iptr++; 3084827bd09bSSatish Balay } 30853fdc5746SBarry Smith ierr = MPI_Isend(dptr3, *msg_size++ *step, MPIU_SCALAR, *msg_list++, MSGTAG1+my_id, gs->gs_comm, msg_ids_out++);CHKERRQ(ierr); 3086827bd09bSSatish Balay } 3087827bd09bSSatish Balay 3088827bd09bSSatish Balay /* tree */ 3089827bd09bSSatish Balay if (gs->max_left_over) 3090827bd09bSSatish Balay {gs_gop_vec_tree_plus(gs,in_vals,step);} 3091827bd09bSSatish Balay 3092827bd09bSSatish Balay /* process the received data */ 3093827bd09bSSatish Balay msg_nodes=nodes; 3094a501084fSBarry Smith while ((iptr = *nodes++)){ 3095a501084fSBarry Smith PetscScalar d1 = 1.0; 3096827bd09bSSatish Balay /* Should I check the return value of MPI_Wait() or status? */ 3097827bd09bSSatish Balay /* Can this loop be replaced by a call to MPI_Waitall()? */ 30983fdc5746SBarry Smith ierr = MPI_Wait(ids_in++, &status);CHKERRQ(ierr); 3099a501084fSBarry Smith while (*iptr >= 0) { 310071044d3cSBarry Smith BLASaxpy_(&step,&d1,in2,&i1,dptr1 + *iptr*step,&i1); 3101827bd09bSSatish Balay in2+=step; 3102827bd09bSSatish Balay iptr++; 3103827bd09bSSatish Balay } 3104827bd09bSSatish Balay } 3105827bd09bSSatish Balay 3106827bd09bSSatish Balay /* replace vals */ 3107827bd09bSSatish Balay while (*pw >= 0) 3108827bd09bSSatish Balay { 3109827bd09bSSatish Balay rvec_copy(in_vals + *pw*step,dptr1,step); 3110827bd09bSSatish Balay dptr1+=step; 3111827bd09bSSatish Balay pw++; 3112827bd09bSSatish Balay } 3113827bd09bSSatish Balay 3114827bd09bSSatish Balay /* clear isend message handles */ 3115827bd09bSSatish Balay /* This changed for clarity though it could be the same */ 3116827bd09bSSatish Balay while (*msg_nodes++) 3117827bd09bSSatish Balay /* Should I check the return value of MPI_Wait() or status? */ 3118827bd09bSSatish Balay /* Can this loop be replaced by a call to MPI_Waitall()? */ 31193fdc5746SBarry Smith {ierr = MPI_Wait(ids_out++, &status);CHKERRQ(ierr);} 3120827bd09bSSatish Balay 31213fdc5746SBarry Smith PetscFunctionReturn(0); 3122827bd09bSSatish Balay } 3123827bd09bSSatish Balay 31247b1ae94cSBarry Smith /******************************************************************************/ 312552f87cdaSBarry Smith static PetscErrorCode gs_gop_vec_tree_plus( gs_id *gs, PetscScalar *vals, PetscInt step) 3126827bd09bSSatish Balay { 312752f87cdaSBarry Smith PetscInt size, *in, *out; 3128a501084fSBarry Smith PetscScalar *buf, *work; 312952f87cdaSBarry Smith PetscInt op[] = {GL_ADD,0}; 3130a501084fSBarry Smith PetscBLASInt i1 = 1; 3131827bd09bSSatish Balay 31323fdc5746SBarry Smith PetscFunctionBegin; 3133827bd09bSSatish Balay /* copy over to local variables */ 3134827bd09bSSatish Balay in = gs->tree_map_in; 3135827bd09bSSatish Balay out = gs->tree_map_out; 3136827bd09bSSatish Balay buf = gs->tree_buf; 3137827bd09bSSatish Balay work = gs->tree_work; 3138827bd09bSSatish Balay size = gs->tree_nel*step; 3139827bd09bSSatish Balay 3140827bd09bSSatish Balay /* zero out collection buffer */ 3141827bd09bSSatish Balay rvec_zero(buf,size); 3142827bd09bSSatish Balay 3143827bd09bSSatish Balay 3144827bd09bSSatish Balay /* copy over my contributions */ 3145827bd09bSSatish Balay while (*in >= 0) 3146827bd09bSSatish Balay { 314771044d3cSBarry Smith BLAScopy_(&step,vals + *in++*step,&i1,buf + *out++*step,&i1); 3148827bd09bSSatish Balay } 3149827bd09bSSatish Balay 3150827bd09bSSatish Balay /* perform fan in/out on full buffer */ 3151827bd09bSSatish Balay /* must change grop to handle the blas */ 3152827bd09bSSatish Balay grop(buf,work,size,op); 3153827bd09bSSatish Balay 3154827bd09bSSatish Balay /* reset */ 3155827bd09bSSatish Balay in = gs->tree_map_in; 3156827bd09bSSatish Balay out = gs->tree_map_out; 3157827bd09bSSatish Balay 3158827bd09bSSatish Balay /* get the portion of the results I need */ 3159827bd09bSSatish Balay while (*in >= 0) 3160827bd09bSSatish Balay { 316171044d3cSBarry Smith BLAScopy_(&step,buf + *out++*step,&i1,vals + *in++*step,&i1); 3162827bd09bSSatish Balay } 31633fdc5746SBarry Smith PetscFunctionReturn(0); 3164827bd09bSSatish Balay } 3165827bd09bSSatish Balay 31667b1ae94cSBarry Smith /******************************************************************************/ 316752f87cdaSBarry Smith PetscErrorCode gs_gop_hc( gs_id *gs, PetscScalar *vals, const char *op, PetscInt dim) 3168827bd09bSSatish Balay { 3169*d1528f56SBarry Smith PetscErrorCode ierr; 3170*d1528f56SBarry Smith 31713fdc5746SBarry Smith PetscFunctionBegin; 3172827bd09bSSatish Balay switch (*op) { 3173827bd09bSSatish Balay case '+': 3174827bd09bSSatish Balay gs_gop_plus_hc(gs,vals,dim); 3175827bd09bSSatish Balay break; 3176827bd09bSSatish Balay default: 3177*d1528f56SBarry Smith ierr = PetscInfo1(0,"gs_gop_hc() :: %c is not a valid op",op[0]); 3178*d1528f56SBarry Smith ierr = PetscInfo(0,"gs_gop_hc() :: default :: plus\n"); 3179827bd09bSSatish Balay gs_gop_plus_hc(gs,vals,dim); 3180827bd09bSSatish Balay break; 3181827bd09bSSatish Balay } 31823fdc5746SBarry Smith PetscFunctionReturn(0); 3183827bd09bSSatish Balay } 3184827bd09bSSatish Balay 31857b1ae94cSBarry Smith /******************************************************************************/ 318652f87cdaSBarry Smith static PetscErrorCode gs_gop_plus_hc( gs_id *gs, PetscScalar *vals, PetscInt dim) 3187827bd09bSSatish Balay { 31883fdc5746SBarry Smith PetscFunctionBegin; 3189827bd09bSSatish Balay /* if there's nothing to do return */ 3190827bd09bSSatish Balay if (dim<=0) 31913fdc5746SBarry Smith { PetscFunctionReturn(0);} 3192827bd09bSSatish Balay 3193827bd09bSSatish Balay /* can't do more dimensions then exist */ 319439945688SSatish Balay dim = PetscMin(dim,i_log2_num_nodes); 3195827bd09bSSatish Balay 3196827bd09bSSatish Balay /* local only operations!!! */ 3197827bd09bSSatish Balay if (gs->num_local) 3198827bd09bSSatish Balay {gs_gop_local_plus(gs,vals);} 3199827bd09bSSatish Balay 3200827bd09bSSatish Balay /* if intersection tree/pairwise and local isn't empty */ 3201827bd09bSSatish Balay if (gs->num_local_gop) 3202827bd09bSSatish Balay { 3203827bd09bSSatish Balay gs_gop_local_in_plus(gs,vals); 3204827bd09bSSatish Balay 3205827bd09bSSatish Balay /* pairwise will do tree inside ... */ 3206827bd09bSSatish Balay if (gs->num_pairs) 3207827bd09bSSatish Balay {gs_gop_pairwise_plus_hc(gs,vals,dim);} 3208827bd09bSSatish Balay 3209827bd09bSSatish Balay /* tree only */ 3210827bd09bSSatish Balay else if (gs->max_left_over) 3211827bd09bSSatish Balay {gs_gop_tree_plus_hc(gs,vals,dim);} 3212827bd09bSSatish Balay 3213827bd09bSSatish Balay gs_gop_local_out(gs,vals); 3214827bd09bSSatish Balay } 3215827bd09bSSatish Balay /* if intersection tree/pairwise and local is empty */ 3216827bd09bSSatish Balay else 3217827bd09bSSatish Balay { 3218827bd09bSSatish Balay /* pairwise will do tree inside */ 3219827bd09bSSatish Balay if (gs->num_pairs) 3220827bd09bSSatish Balay {gs_gop_pairwise_plus_hc(gs,vals,dim);} 3221827bd09bSSatish Balay 3222827bd09bSSatish Balay /* tree */ 3223827bd09bSSatish Balay else if (gs->max_left_over) 3224827bd09bSSatish Balay {gs_gop_tree_plus_hc(gs,vals,dim);} 3225827bd09bSSatish Balay } 32263fdc5746SBarry Smith PetscFunctionReturn(0); 3227827bd09bSSatish Balay } 3228827bd09bSSatish Balay 32297b1ae94cSBarry Smith /******************************************************************************/ 323052f87cdaSBarry Smith static PetscErrorCode gs_gop_pairwise_plus_hc( gs_id *gs, PetscScalar *in_vals, PetscInt dim) 3231827bd09bSSatish Balay { 3232a501084fSBarry Smith PetscScalar *dptr1, *dptr2, *dptr3, *in1, *in2; 323352f87cdaSBarry Smith PetscInt *iptr, *msg_list, *msg_size, **msg_nodes; 323452f87cdaSBarry Smith PetscInt *pw, *list, *size, **nodes; 3235827bd09bSSatish Balay MPI_Request *msg_ids_in, *msg_ids_out, *ids_in, *ids_out; 3236827bd09bSSatish Balay MPI_Status status; 323752f87cdaSBarry Smith PetscInt i, mask=1; 32383fdc5746SBarry Smith PetscErrorCode ierr; 3239827bd09bSSatish Balay 32403fdc5746SBarry Smith PetscFunctionBegin; 3241827bd09bSSatish Balay for (i=1; i<dim; i++) 3242827bd09bSSatish Balay {mask<<=1; mask++;} 3243827bd09bSSatish Balay 3244827bd09bSSatish Balay 3245a501084fSBarry Smith /* strip and load s */ 3246827bd09bSSatish Balay msg_list =list = gs->pair_list; 3247827bd09bSSatish Balay msg_size =size = gs->msg_sizes; 3248827bd09bSSatish Balay msg_nodes=nodes = gs->node_list; 3249827bd09bSSatish Balay iptr=pw = gs->pw_elm_list; 3250827bd09bSSatish Balay dptr1=dptr3 = gs->pw_vals; 3251827bd09bSSatish Balay msg_ids_in = ids_in = gs->msg_ids_in; 3252827bd09bSSatish Balay msg_ids_out = ids_out = gs->msg_ids_out; 3253827bd09bSSatish Balay dptr2 = gs->out; 3254827bd09bSSatish Balay in1=in2 = gs->in; 3255827bd09bSSatish Balay 3256827bd09bSSatish Balay /* post the receives */ 3257827bd09bSSatish Balay /* msg_nodes=nodes; */ 3258827bd09bSSatish Balay do 3259827bd09bSSatish Balay { 3260827bd09bSSatish Balay /* Should MPI_ANY_SOURCE be replaced by *list ? In that case do the 3261827bd09bSSatish Balay second one *list and do list++ afterwards */ 3262827bd09bSSatish Balay if ((my_id|mask)==(*list|mask)) 3263827bd09bSSatish Balay { 32643fdc5746SBarry Smith ierr = MPI_Irecv(in1, *size, MPIU_SCALAR, MPI_ANY_SOURCE, MSGTAG1 + *list++, gs->gs_comm, msg_ids_in++);CHKERRQ(ierr); 3265827bd09bSSatish Balay in1 += *size++; 3266827bd09bSSatish Balay } 3267827bd09bSSatish Balay else 3268827bd09bSSatish Balay {list++; size++;} 3269827bd09bSSatish Balay } 3270827bd09bSSatish Balay while (*++msg_nodes); 3271827bd09bSSatish Balay 3272827bd09bSSatish Balay /* load gs values into in out gs buffers */ 3273827bd09bSSatish Balay while (*iptr >= 0) 3274827bd09bSSatish Balay {*dptr3++ = *(in_vals + *iptr++);} 3275827bd09bSSatish Balay 3276827bd09bSSatish Balay /* load out buffers and post the sends */ 3277827bd09bSSatish Balay msg_nodes=nodes; 3278827bd09bSSatish Balay list = msg_list; 3279827bd09bSSatish Balay while ((iptr = *msg_nodes++)) 3280827bd09bSSatish Balay { 3281827bd09bSSatish Balay if ((my_id|mask)==(*list|mask)) 3282827bd09bSSatish Balay { 3283827bd09bSSatish Balay dptr3 = dptr2; 3284827bd09bSSatish Balay while (*iptr >= 0) 3285827bd09bSSatish Balay {*dptr2++ = *(dptr1 + *iptr++);} 3286827bd09bSSatish Balay /* CHECK PERSISTENT COMMS MODE FOR ALL THIS STUFF */ 3287827bd09bSSatish Balay /* is msg_ids_out++ correct? */ 32883fdc5746SBarry Smith ierr = MPI_Isend(dptr3, *msg_size++, MPIU_SCALAR, *list++, MSGTAG1+my_id, gs->gs_comm, msg_ids_out++);CHKERRQ(ierr); 3289827bd09bSSatish Balay } 3290827bd09bSSatish Balay else 3291827bd09bSSatish Balay {list++; msg_size++;} 3292827bd09bSSatish Balay } 3293827bd09bSSatish Balay 3294827bd09bSSatish Balay /* do the tree while we're waiting */ 3295827bd09bSSatish Balay if (gs->max_left_over) 3296827bd09bSSatish Balay {gs_gop_tree_plus_hc(gs,in_vals,dim);} 3297827bd09bSSatish Balay 3298827bd09bSSatish Balay /* process the received data */ 3299827bd09bSSatish Balay msg_nodes=nodes; 3300827bd09bSSatish Balay list = msg_list; 3301827bd09bSSatish Balay while ((iptr = *nodes++)) 3302827bd09bSSatish Balay { 3303827bd09bSSatish Balay if ((my_id|mask)==(*list|mask)) 3304827bd09bSSatish Balay { 3305827bd09bSSatish Balay /* Should I check the return value of MPI_Wait() or status? */ 3306827bd09bSSatish Balay /* Can this loop be replaced by a call to MPI_Waitall()? */ 33073fdc5746SBarry Smith ierr = MPI_Wait(ids_in++, &status);CHKERRQ(ierr); 3308827bd09bSSatish Balay while (*iptr >= 0) 3309827bd09bSSatish Balay {*(dptr1 + *iptr++) += *in2++;} 3310827bd09bSSatish Balay } 3311827bd09bSSatish Balay list++; 3312827bd09bSSatish Balay } 3313827bd09bSSatish Balay 3314827bd09bSSatish Balay /* replace vals */ 3315827bd09bSSatish Balay while (*pw >= 0) 3316827bd09bSSatish Balay {*(in_vals + *pw++) = *dptr1++;} 3317827bd09bSSatish Balay 3318827bd09bSSatish Balay /* clear isend message handles */ 3319827bd09bSSatish Balay /* This changed for clarity though it could be the same */ 3320827bd09bSSatish Balay while (*msg_nodes++) 3321827bd09bSSatish Balay { 3322827bd09bSSatish Balay if ((my_id|mask)==(*msg_list|mask)) 3323827bd09bSSatish Balay { 3324827bd09bSSatish Balay /* Should I check the return value of MPI_Wait() or status? */ 3325827bd09bSSatish Balay /* Can this loop be replaced by a call to MPI_Waitall()? */ 33263fdc5746SBarry Smith ierr = MPI_Wait(ids_out++, &status);CHKERRQ(ierr); 3327827bd09bSSatish Balay } 3328827bd09bSSatish Balay msg_list++; 3329827bd09bSSatish Balay } 3330827bd09bSSatish Balay 33313fdc5746SBarry Smith PetscFunctionReturn(0); 3332827bd09bSSatish Balay } 3333827bd09bSSatish Balay 33347b1ae94cSBarry Smith /******************************************************************************/ 333552f87cdaSBarry Smith static PetscErrorCode gs_gop_tree_plus_hc(gs_id *gs, PetscScalar *vals, PetscInt dim) 3336827bd09bSSatish Balay { 333752f87cdaSBarry Smith PetscInt size; 333852f87cdaSBarry Smith PetscInt *in, *out; 3339a501084fSBarry Smith PetscScalar *buf, *work; 334052f87cdaSBarry Smith PetscInt op[] = {GL_ADD,0}; 3341827bd09bSSatish Balay 33423fdc5746SBarry Smith PetscFunctionBegin; 3343827bd09bSSatish Balay in = gs->tree_map_in; 3344827bd09bSSatish Balay out = gs->tree_map_out; 3345827bd09bSSatish Balay buf = gs->tree_buf; 3346827bd09bSSatish Balay work = gs->tree_work; 3347827bd09bSSatish Balay size = gs->tree_nel; 3348827bd09bSSatish Balay 3349827bd09bSSatish Balay rvec_zero(buf,size); 3350827bd09bSSatish Balay 3351827bd09bSSatish Balay while (*in >= 0) 3352827bd09bSSatish Balay {*(buf + *out++) = *(vals + *in++);} 3353827bd09bSSatish Balay 3354827bd09bSSatish Balay in = gs->tree_map_in; 3355827bd09bSSatish Balay out = gs->tree_map_out; 3356827bd09bSSatish Balay 3357827bd09bSSatish Balay grop_hc(buf,work,size,op,dim); 3358827bd09bSSatish Balay 3359827bd09bSSatish Balay while (*in >= 0) 3360827bd09bSSatish Balay {*(vals + *in++) = *(buf + *out++);} 33613fdc5746SBarry Smith PetscFunctionReturn(0); 3362827bd09bSSatish Balay } 3363827bd09bSSatish Balay 3364827bd09bSSatish Balay 3365827bd09bSSatish Balay 3366