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 { 38827bd09bSSatish Balay int id; 39827bd09bSSatish Balay int nel_min; 40827bd09bSSatish Balay int nel_max; 41827bd09bSSatish Balay int nel_sum; 42827bd09bSSatish Balay int negl; 43827bd09bSSatish Balay int gl_max; 44827bd09bSSatish Balay int gl_min; 45827bd09bSSatish Balay int repeats; 46827bd09bSSatish Balay int ordered; 47827bd09bSSatish Balay int positive; 48a501084fSBarry Smith PetscScalar *vals; 49827bd09bSSatish Balay 50827bd09bSSatish Balay /* bit mask info */ 51827bd09bSSatish Balay int *my_proc_mask; 52827bd09bSSatish Balay int mask_sz; 53827bd09bSSatish Balay int *ngh_buf; 54827bd09bSSatish Balay int ngh_buf_sz; 55827bd09bSSatish Balay int *nghs; 56827bd09bSSatish Balay int num_nghs; 57827bd09bSSatish Balay int max_nghs; 58827bd09bSSatish Balay int *pw_nghs; 59827bd09bSSatish Balay int num_pw_nghs; 60827bd09bSSatish Balay int *tree_nghs; 61827bd09bSSatish Balay int num_tree_nghs; 62827bd09bSSatish Balay 63827bd09bSSatish Balay int num_loads; 64827bd09bSSatish Balay 65827bd09bSSatish Balay /* repeats == true -> local info */ 66827bd09bSSatish Balay int nel; /* number of unique elememts */ 67827bd09bSSatish Balay int *elms; /* of size nel */ 68827bd09bSSatish Balay int nel_total; 69827bd09bSSatish Balay int *local_elms; /* of size nel_total */ 70827bd09bSSatish Balay int *companion; /* of size nel_total */ 71827bd09bSSatish Balay 72827bd09bSSatish Balay /* local info */ 73827bd09bSSatish Balay int num_local_total; 74827bd09bSSatish Balay int local_strength; 75827bd09bSSatish Balay int num_local; 76827bd09bSSatish Balay int *num_local_reduce; 77827bd09bSSatish Balay int **local_reduce; 78827bd09bSSatish Balay int num_local_gop; 79827bd09bSSatish Balay int *num_gop_local_reduce; 80827bd09bSSatish Balay int **gop_local_reduce; 81827bd09bSSatish Balay 82827bd09bSSatish Balay /* pairwise info */ 83827bd09bSSatish Balay int level; 84827bd09bSSatish Balay int num_pairs; 85827bd09bSSatish Balay int max_pairs; 86827bd09bSSatish Balay int loc_node_pairs; 87827bd09bSSatish Balay int max_node_pairs; 88827bd09bSSatish Balay int min_node_pairs; 89827bd09bSSatish Balay int avg_node_pairs; 90827bd09bSSatish Balay int *pair_list; 91827bd09bSSatish Balay int *msg_sizes; 92827bd09bSSatish Balay int **node_list; 93827bd09bSSatish Balay int len_pw_list; 94827bd09bSSatish Balay int *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; 102827bd09bSSatish Balay int msg_total; 103827bd09bSSatish Balay 104827bd09bSSatish Balay /* tree - crystal accumulator info */ 105827bd09bSSatish Balay int max_left_over; 106827bd09bSSatish Balay int *pre; 107827bd09bSSatish Balay int *in_num; 108827bd09bSSatish Balay int *out_num; 109827bd09bSSatish Balay int **in_list; 110827bd09bSSatish Balay int **out_list; 111827bd09bSSatish Balay 112827bd09bSSatish Balay /* new tree work*/ 113827bd09bSSatish Balay int tree_nel; 114827bd09bSSatish Balay int *tree_elms; 115a501084fSBarry Smith PetscScalar *tree_buf; 116a501084fSBarry Smith PetscScalar *tree_work; 117827bd09bSSatish Balay 118827bd09bSSatish Balay int tree_map_sz; 119827bd09bSSatish Balay int *tree_map_in; 120827bd09bSSatish Balay int *tree_map_out; 121827bd09bSSatish Balay 122827bd09bSSatish Balay /* current memory status */ 123827bd09bSSatish Balay int gl_bss_min; 124827bd09bSSatish Balay int gl_perm_min; 125827bd09bSSatish Balay 126827bd09bSSatish Balay /* max segment size for gs_gop_vec() */ 127827bd09bSSatish Balay int 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 134827bd09bSSatish Balay static gs_id *gsi_check_args(int *elms, int nel, int 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 */ 1443fdc5746SBarry Smith static PetscErrorCode gs_gop_vec_local_out(gs_id *gs, PetscScalar *vals, int step); 145827bd09bSSatish Balay 1463fdc5746SBarry Smith static PetscErrorCode gs_gop_vec_plus(gs_id *gs, PetscScalar *in_vals, int step); 1473fdc5746SBarry Smith static PetscErrorCode gs_gop_vec_pairwise_plus(gs_id *gs, PetscScalar *in_vals, int step); 1483fdc5746SBarry Smith static PetscErrorCode gs_gop_vec_local_plus(gs_id *gs, PetscScalar *vals, int step); 1493fdc5746SBarry Smith static PetscErrorCode gs_gop_vec_local_in_plus(gs_id *gs, PetscScalar *vals, int step); 1503fdc5746SBarry Smith static PetscErrorCode gs_gop_vec_tree_plus(gs_id *gs, PetscScalar *vals, int 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 1593fdc5746SBarry Smith static PetscErrorCode gs_gop_plus_hc(gs_id *gs, PetscScalar *in_vals, int dim); 1603fdc5746SBarry Smith static PetscErrorCode gs_gop_pairwise_plus_hc(gs_id *gs, PetscScalar *in_vals, int dim); 1613fdc5746SBarry Smith static PetscErrorCode gs_gop_tree_plus_hc(gs_id *gs, PetscScalar *vals, int 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 209827bd09bSSatish Balay static int num_gs_ids = 0; 210827bd09bSSatish Balay 211827bd09bSSatish Balay /* should make this dynamic ... later */ 212827bd09bSSatish Balay static int msg_buf=MAX_MSG_BUF; 213827bd09bSSatish Balay static int vec_sz=GS_VEC_SZ; 214827bd09bSSatish Balay static int *tree_buf=NULL; 215827bd09bSSatish Balay static int tree_buf_sz=0; 216827bd09bSSatish Balay static int 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 ******************************************************************************/ 2273fdc5746SBarry Smith PetscErrorCode gs_init_vec_sz(int 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 ******************************************************************************/ 2423fdc5746SBarry Smith PetscErrorCode gs_init_msg_buf_sz(int 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 ******************************************************************************/ 2600924e98cSBarry Smith gs_id *gs_init( int *elms, int nel, int 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 ******************************************************************************/ 3220924e98cSBarry Smith static gs_id * gsi_check_args(int *in_elms, int nel, int level) 323827bd09bSSatish Balay { 324a501084fSBarry Smith int i, j, k, t2; 325827bd09bSSatish Balay int *companion, *elms, *unique, *iptr; 326827bd09bSSatish Balay int num_local=0, *num_to_reduce, **local_reduce; 327827bd09bSSatish Balay int oprs[] = {NON_UNIFORM,GL_MIN,GL_MAX,GL_ADD,GL_MIN,GL_MAX,GL_MIN,GL_B_AND}; 328827bd09bSSatish Balay int vals[sizeof(oprs)/sizeof(oprs[0])-1]; 329827bd09bSSatish Balay int work[sizeof(oprs)/sizeof(oprs[0])-1]; 330827bd09bSSatish Balay gs_id *gs; 331827bd09bSSatish Balay 332827bd09bSSatish Balay 333827bd09bSSatish Balay 334827bd09bSSatish Balay if (!in_elms) 335827bd09bSSatish Balay {error_msg_fatal("elms point to nothing!!!\n");} 336827bd09bSSatish Balay 337827bd09bSSatish Balay if (nel<0) 338827bd09bSSatish Balay {error_msg_fatal("can't have fewer than 0 elms!!!\n");} 339827bd09bSSatish Balay 340827bd09bSSatish Balay if (nel==0) 341827bd09bSSatish Balay {error_msg_warning("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 */ 357a501084fSBarry Smith elms = (int*) malloc((nel+1)*sizeof(PetscInt)); 358a501084fSBarry Smith companion = (int*) malloc(nel*sizeof(PetscInt)); 359*1d7d0905SBarry 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) 367827bd09bSSatish Balay {error_msg_fatal("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 { 383827bd09bSSatish Balay error_msg_warning("gsi_check_args() :: elm list *not* sorted!\n"); 384827bd09bSSatish Balay SMI_sort((void*)elms, (void*)companion, nel, SORT_INTEGER); 385827bd09bSSatish Balay } 386827bd09bSSatish Balay else 387827bd09bSSatish Balay {error_msg_warning("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; 413a501084fSBarry Smith gs->local_reduce=local_reduce=(int **)malloc(num_local*sizeof(PetscInt*)); 414a501084fSBarry Smith gs->num_local_reduce=num_to_reduce=(int*) malloc(num_local*sizeof(PetscInt)); 415827bd09bSSatish Balay 416a501084fSBarry Smith unique = (int*) 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++; 435a501084fSBarry Smith iptr = local_reduce[num_local++] = (int*)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 { 461*1d7d0905SBarry Smith vals[3] = unique[0]; 462*1d7d0905SBarry Smith vals[4] = unique[gs->nel-1]; 463827bd09bSSatish Balay } 464827bd09bSSatish Balay else 465827bd09bSSatish Balay { 466*1d7d0905SBarry Smith vals[3] = INT_MAX; 467*1d7d0905SBarry 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) 478827bd09bSSatish Balay {error_msg_fatal("gsi_check_args() :: system not semi-pos def ::%d\n",vals[3]);} 479827bd09bSSatish Balay 480827bd09bSSatish Balay if (vals[4]==INT_MAX) 481827bd09bSSatish Balay {error_msg_fatal("gsi_check_args() :: system ub too large ::%d!\n",vals[4]);} 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) 491827bd09bSSatish Balay {error_msg_fatal("gsi_check_args() :: system empty or neg :: %d\n",gs->negl);} 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 { 516a501084fSBarry Smith int i, nel, *elms; 517827bd09bSSatish Balay int t1; 518827bd09bSSatish Balay int **reduce; 519827bd09bSSatish Balay int *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) 554827bd09bSSatish Balay {error_msg_fatal("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) 582827bd09bSSatish Balay {error_msg_fatal("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 ******************************************************************************/ 6230924e98cSBarry Smith static PetscErrorCode place_in_tree( int elm) 624827bd09bSSatish Balay { 625a501084fSBarry Smith int *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; 635a501084fSBarry Smith tree_buf = (int*)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; 642a501084fSBarry Smith tree_buf = (int*)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 { 664a501084fSBarry Smith int i, j, npw=0, ntree_map=0; 665827bd09bSSatish Balay int p_mask_size, ngh_buf_size, buf_size; 666827bd09bSSatish Balay int *p_mask, *sh_proc_mask, *pw_sh_proc_mask; 667827bd09bSSatish Balay int *ngh_buf, *buf1, *buf2; 668827bd09bSSatish Balay int offset, per_load, num_loads, or_ct, start, end; 669827bd09bSSatish Balay int *ptr1, *ptr2, i_start, negl, nel, *elms; 670827bd09bSSatish Balay int oper=GL_B_OR; 671827bd09bSSatish Balay int *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 */ 680a501084fSBarry Smith p_mask = (int*) 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 */ 684a501084fSBarry Smith gs->nghs = sh_proc_mask = (int*) malloc(p_mask_size); 685a501084fSBarry Smith gs->pw_nghs = pw_sh_proc_mask = (int*) malloc(p_mask_size); 686827bd09bSSatish Balay gs->ngh_buf_sz = ngh_buf_size = p_mask_size*nel; 687a501084fSBarry Smith t_mask = (int*) malloc(p_mask_size); 688a501084fSBarry Smith gs->ngh_buf = ngh_buf = (int*) 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) 706827bd09bSSatish Balay {error_msg_fatal("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 */ 709a501084fSBarry Smith buf1 = (int*) 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 /* queue the tree elements for now */ 739827bd09bSSatish Balay /* elms_q = new_queue(); */ 740827bd09bSSatish Balay 741827bd09bSSatish Balay /* can also queue tree info for pruned or forest implememtation */ 742827bd09bSSatish Balay /* mask_q = new_queue(); */ 743827bd09bSSatish Balay 744827bd09bSSatish Balay /* ok do it */ 745827bd09bSSatish Balay for (ptr1=ngh_buf,ptr2=elms,end=gs->gl_min,or_ct=i=0; or_ct<num_loads; or_ct++) 746827bd09bSSatish Balay { 747827bd09bSSatish Balay /* identity for bitwise or is 000...000 */ 748827bd09bSSatish Balay ivec_zero(buf1,buf_size); 749827bd09bSSatish Balay 750827bd09bSSatish Balay /* load msg buffer */ 751827bd09bSSatish Balay for (start=end,end+=per_load,i_start=i; (offset=*ptr2)<end; i++, ptr2++) 752827bd09bSSatish Balay { 753827bd09bSSatish Balay offset = (offset-start)*p_mask_size; 754827bd09bSSatish Balay ivec_copy(buf1+offset,p_mask,p_mask_size); 755827bd09bSSatish Balay } 756827bd09bSSatish Balay 757827bd09bSSatish Balay /* GLOBAL: pass buffer */ 758827bd09bSSatish Balay giop(buf1,buf2,buf_size,&oper); 759827bd09bSSatish Balay 760827bd09bSSatish Balay 761827bd09bSSatish Balay /* unload buffer into ngh_buf */ 762827bd09bSSatish Balay ptr2=(elms+i_start); 763827bd09bSSatish Balay for(ptr3=buf1,j=start; j<end; ptr3+=p_mask_size,j++) 764827bd09bSSatish Balay { 765827bd09bSSatish Balay /* I own it ... may have to pairwise it */ 766827bd09bSSatish Balay if (j==*ptr2) 767827bd09bSSatish Balay { 768827bd09bSSatish Balay /* do i share it w/anyone? */ 769a501084fSBarry Smith ct1 = ct_bits((char *)ptr3,p_mask_size*sizeof(PetscInt)); 770827bd09bSSatish Balay /* guess not */ 771827bd09bSSatish Balay if (ct1<2) 772827bd09bSSatish Balay {ptr2++; ptr1+=p_mask_size; continue;} 773827bd09bSSatish Balay 774827bd09bSSatish Balay /* i do ... so keep info and turn off my bit */ 775827bd09bSSatish Balay ivec_copy(ptr1,ptr3,p_mask_size); 776827bd09bSSatish Balay ivec_xor(ptr1,p_mask,p_mask_size); 777827bd09bSSatish Balay ivec_or(sh_proc_mask,ptr1,p_mask_size); 778827bd09bSSatish Balay 779827bd09bSSatish Balay /* is it to be done pairwise? */ 780827bd09bSSatish Balay if (--ct1<=level) 781827bd09bSSatish Balay { 782827bd09bSSatish Balay npw++; 783827bd09bSSatish Balay 784827bd09bSSatish Balay /* turn on high bit to indicate pw need to process */ 785827bd09bSSatish Balay *ptr2++ |= TOP_BIT; 786827bd09bSSatish Balay ivec_or(pw_sh_proc_mask,ptr1,p_mask_size); 787827bd09bSSatish Balay ptr1+=p_mask_size; 788827bd09bSSatish Balay continue; 789827bd09bSSatish Balay } 790827bd09bSSatish Balay 791827bd09bSSatish Balay /* get set for next and note that I have a tree contribution */ 792827bd09bSSatish Balay /* could save exact elm index for tree here -> save a search */ 793827bd09bSSatish Balay ptr2++; ptr1+=p_mask_size; ntree_map++; 794827bd09bSSatish Balay } 795827bd09bSSatish Balay /* i don't but still might be involved in tree */ 796827bd09bSSatish Balay else 797827bd09bSSatish Balay { 798827bd09bSSatish Balay 799827bd09bSSatish Balay /* shared by how many? */ 800a501084fSBarry Smith ct1 = ct_bits((char *)ptr3,p_mask_size*sizeof(PetscInt)); 801827bd09bSSatish Balay 802827bd09bSSatish Balay /* none! */ 803827bd09bSSatish Balay if (ct1<2) 804827bd09bSSatish Balay {continue;} 805827bd09bSSatish Balay 806827bd09bSSatish Balay /* is it going to be done pairwise? but not by me of course!*/ 807827bd09bSSatish Balay if (--ct1<=level) 808827bd09bSSatish Balay {continue;} 809827bd09bSSatish Balay } 810827bd09bSSatish Balay /* LATER we're going to have to process it NOW */ 811827bd09bSSatish Balay /* nope ... tree it */ 812827bd09bSSatish Balay place_in_tree(j); 813827bd09bSSatish Balay } 814827bd09bSSatish Balay } 815827bd09bSSatish Balay 816a501084fSBarry Smith free((void*)t_mask); 817a501084fSBarry Smith free((void*)buf1); 818827bd09bSSatish Balay 819827bd09bSSatish Balay gs->len_pw_list=npw; 820a501084fSBarry Smith gs->num_nghs = ct_bits((char *)sh_proc_mask,p_mask_size*sizeof(PetscInt)); 821827bd09bSSatish Balay 822827bd09bSSatish Balay /* expand from bit mask list to int list and save ngh list */ 823a501084fSBarry Smith gs->nghs = (int*) malloc(gs->num_nghs * sizeof(PetscInt)); 824a501084fSBarry Smith bm_to_proc((char *)sh_proc_mask,p_mask_size*sizeof(PetscInt),gs->nghs); 825827bd09bSSatish Balay 826a501084fSBarry Smith gs->num_pw_nghs = ct_bits((char *)pw_sh_proc_mask,p_mask_size*sizeof(PetscInt)); 827827bd09bSSatish Balay 828827bd09bSSatish Balay oper = GL_MAX; 829827bd09bSSatish Balay ct1 = gs->num_nghs; 830827bd09bSSatish Balay giop(&ct1,&ct2,1,&oper); 831827bd09bSSatish Balay gs->max_nghs = ct1; 832827bd09bSSatish Balay 833827bd09bSSatish Balay gs->tree_map_sz = ntree_map; 834827bd09bSSatish Balay gs->max_left_over=ntree; 835827bd09bSSatish Balay 836a501084fSBarry Smith free((void*)p_mask); 837a501084fSBarry Smith free((void*)sh_proc_mask); 8383fdc5746SBarry Smith PetscFunctionReturn(0); 839827bd09bSSatish Balay } 840827bd09bSSatish Balay 841827bd09bSSatish Balay 842827bd09bSSatish Balay 843827bd09bSSatish Balay 844827bd09bSSatish Balay 845827bd09bSSatish Balay /****************************************************************************** 846827bd09bSSatish Balay Function: pairwise_init() 847827bd09bSSatish Balay 848827bd09bSSatish Balay Input : 849827bd09bSSatish Balay Output: 850827bd09bSSatish Balay Return: 851827bd09bSSatish Balay Description: 852827bd09bSSatish Balay 853827bd09bSSatish Balay if an element is shared by fewer that level# of nodes do pairwise exch 854827bd09bSSatish Balay ******************************************************************************/ 8550924e98cSBarry Smith static PetscErrorCode set_pairwise(gs_id *gs) 856827bd09bSSatish Balay { 857a501084fSBarry Smith int i, j; 858827bd09bSSatish Balay int p_mask_size; 859827bd09bSSatish Balay int *p_mask, *sh_proc_mask, *tmp_proc_mask; 860827bd09bSSatish Balay int *ngh_buf, *buf2; 861827bd09bSSatish Balay int offset; 862827bd09bSSatish Balay int *msg_list, *msg_size, **msg_nodes, nprs; 863827bd09bSSatish Balay int *pairwise_elm_list, len_pair_list=0; 864827bd09bSSatish Balay int *iptr, t1, i_start, nel, *elms; 865827bd09bSSatish Balay int ct; 866827bd09bSSatish Balay 867827bd09bSSatish Balay 8683fdc5746SBarry Smith PetscFunctionBegin; 869827bd09bSSatish Balay /* to make life easier */ 870827bd09bSSatish Balay nel = gs->nel; 871827bd09bSSatish Balay elms = gs->elms; 872827bd09bSSatish Balay ngh_buf = gs->ngh_buf; 873827bd09bSSatish Balay sh_proc_mask = gs->pw_nghs; 874827bd09bSSatish Balay 875827bd09bSSatish Balay /* need a few temp masks */ 876827bd09bSSatish Balay p_mask_size = len_bit_mask(num_nodes); 877a501084fSBarry Smith p_mask = (int*) malloc(p_mask_size); 878a501084fSBarry Smith tmp_proc_mask = (int*) malloc(p_mask_size); 879827bd09bSSatish Balay 880827bd09bSSatish Balay /* set mask to my my_id's bit mask */ 881827bd09bSSatish Balay set_bit_mask(p_mask,p_mask_size,my_id); 882827bd09bSSatish Balay 883a501084fSBarry Smith p_mask_size /= sizeof(PetscInt); 884827bd09bSSatish Balay 885827bd09bSSatish Balay len_pair_list=gs->len_pw_list; 886a501084fSBarry Smith gs->pw_elm_list=pairwise_elm_list=(int*)malloc((len_pair_list+1)*sizeof(PetscInt)); 887827bd09bSSatish Balay 888827bd09bSSatish Balay /* how many processors (nghs) do we have to exchange with? */ 889a501084fSBarry Smith nprs=gs->num_pairs=ct_bits((char *)sh_proc_mask,p_mask_size*sizeof(PetscInt)); 890827bd09bSSatish Balay 891827bd09bSSatish Balay 892827bd09bSSatish Balay /* allocate space for gs_gop() info */ 893a501084fSBarry Smith gs->pair_list = msg_list = (int*) malloc(sizeof(PetscInt)*nprs); 894a501084fSBarry Smith gs->msg_sizes = msg_size = (int*) malloc(sizeof(PetscInt)*nprs); 895a501084fSBarry Smith gs->node_list = msg_nodes = (int **) malloc(sizeof(PetscInt*)*(nprs+1)); 896827bd09bSSatish Balay 897827bd09bSSatish Balay /* init msg_size list */ 898827bd09bSSatish Balay ivec_zero(msg_size,nprs); 899827bd09bSSatish Balay 900827bd09bSSatish Balay /* expand from bit mask list to int list */ 901a501084fSBarry Smith bm_to_proc((char *)sh_proc_mask,p_mask_size*sizeof(PetscInt),msg_list); 902827bd09bSSatish Balay 903827bd09bSSatish Balay /* keep list of elements being handled pairwise */ 904827bd09bSSatish Balay for (i=j=0;i<nel;i++) 905827bd09bSSatish Balay { 906827bd09bSSatish Balay if (elms[i] & TOP_BIT) 907827bd09bSSatish Balay {elms[i] ^= TOP_BIT; pairwise_elm_list[j++] = i;} 908827bd09bSSatish Balay } 909827bd09bSSatish Balay pairwise_elm_list[j] = -1; 910827bd09bSSatish Balay 911a501084fSBarry Smith gs->msg_ids_out = (MPI_Request *) malloc(sizeof(MPI_Request)*(nprs+1)); 912827bd09bSSatish Balay gs->msg_ids_out[nprs] = MPI_REQUEST_NULL; 913a501084fSBarry Smith gs->msg_ids_in = (MPI_Request *) malloc(sizeof(MPI_Request)*(nprs+1)); 914827bd09bSSatish Balay gs->msg_ids_in[nprs] = MPI_REQUEST_NULL; 915a501084fSBarry Smith gs->pw_vals = (PetscScalar *) malloc(sizeof(PetscScalar)*len_pair_list*vec_sz); 916827bd09bSSatish Balay 917827bd09bSSatish Balay /* find who goes to each processor */ 918827bd09bSSatish Balay for (i_start=i=0;i<nprs;i++) 919827bd09bSSatish Balay { 920827bd09bSSatish Balay /* processor i's mask */ 921a501084fSBarry Smith set_bit_mask(p_mask,p_mask_size*sizeof(PetscInt),msg_list[i]); 922827bd09bSSatish Balay 923827bd09bSSatish Balay /* det # going to processor i */ 924827bd09bSSatish Balay for (ct=j=0;j<len_pair_list;j++) 925827bd09bSSatish Balay { 926827bd09bSSatish Balay buf2 = ngh_buf+(pairwise_elm_list[j]*p_mask_size); 927827bd09bSSatish Balay ivec_and3(tmp_proc_mask,p_mask,buf2,p_mask_size); 928a501084fSBarry Smith if (ct_bits((char *)tmp_proc_mask,p_mask_size*sizeof(PetscInt))) 929827bd09bSSatish Balay {ct++;} 930827bd09bSSatish Balay } 931827bd09bSSatish Balay msg_size[i] = ct; 93239945688SSatish Balay i_start = PetscMax(i_start,ct); 933827bd09bSSatish Balay 934827bd09bSSatish Balay /*space to hold nodes in message to first neighbor */ 935a501084fSBarry Smith msg_nodes[i] = iptr = (int*) malloc(sizeof(PetscInt)*(ct+1)); 936827bd09bSSatish Balay 937827bd09bSSatish Balay for (j=0;j<len_pair_list;j++) 938827bd09bSSatish Balay { 939827bd09bSSatish Balay buf2 = ngh_buf+(pairwise_elm_list[j]*p_mask_size); 940827bd09bSSatish Balay ivec_and3(tmp_proc_mask,p_mask,buf2,p_mask_size); 941a501084fSBarry Smith if (ct_bits((char *)tmp_proc_mask,p_mask_size*sizeof(PetscInt))) 942827bd09bSSatish Balay {*iptr++ = j;} 943827bd09bSSatish Balay } 944827bd09bSSatish Balay *iptr = -1; 945827bd09bSSatish Balay } 946827bd09bSSatish Balay msg_nodes[nprs] = NULL; 947827bd09bSSatish Balay 948827bd09bSSatish Balay j=gs->loc_node_pairs=i_start; 949827bd09bSSatish Balay t1 = GL_MAX; 950827bd09bSSatish Balay giop(&i_start,&offset,1,&t1); 951827bd09bSSatish Balay gs->max_node_pairs = i_start; 952827bd09bSSatish Balay 953827bd09bSSatish Balay i_start=j; 954827bd09bSSatish Balay t1 = GL_MIN; 955827bd09bSSatish Balay giop(&i_start,&offset,1,&t1); 956827bd09bSSatish Balay gs->min_node_pairs = i_start; 957827bd09bSSatish Balay 958827bd09bSSatish Balay i_start=j; 959827bd09bSSatish Balay t1 = GL_ADD; 960827bd09bSSatish Balay giop(&i_start,&offset,1,&t1); 961827bd09bSSatish Balay gs->avg_node_pairs = i_start/num_nodes + 1; 962827bd09bSSatish Balay 963827bd09bSSatish Balay i_start=nprs; 964827bd09bSSatish Balay t1 = GL_MAX; 965827bd09bSSatish Balay giop(&i_start,&offset,1,&t1); 966827bd09bSSatish Balay gs->max_pairs = i_start; 967827bd09bSSatish Balay 968827bd09bSSatish Balay 969827bd09bSSatish Balay /* remap pairwise in tail of gsi_via_bit_mask() */ 970827bd09bSSatish Balay gs->msg_total = ivec_sum(gs->msg_sizes,nprs); 971a501084fSBarry Smith gs->out = (PetscScalar *) malloc(sizeof(PetscScalar)*gs->msg_total*vec_sz); 972a501084fSBarry Smith gs->in = (PetscScalar *) malloc(sizeof(PetscScalar)*gs->msg_total*vec_sz); 973827bd09bSSatish Balay 974827bd09bSSatish Balay /* reset malloc pool */ 975a501084fSBarry Smith free((void*)p_mask); 976a501084fSBarry Smith free((void*)tmp_proc_mask); 9773fdc5746SBarry Smith PetscFunctionReturn(0); 978827bd09bSSatish Balay } 979827bd09bSSatish Balay 980827bd09bSSatish Balay 981827bd09bSSatish Balay 982827bd09bSSatish Balay /****************************************************************************** 983827bd09bSSatish Balay Function: set_tree() 984827bd09bSSatish Balay 985827bd09bSSatish Balay Input : 986827bd09bSSatish Balay Output: 987827bd09bSSatish Balay Return: 988827bd09bSSatish Balay Description: 989827bd09bSSatish Balay 990827bd09bSSatish Balay to do pruned tree just save ngh buf copy for each one and decode here! 991827bd09bSSatish Balay ******************************************************************************/ 9920924e98cSBarry Smith static PetscErrorCode set_tree(gs_id *gs) 993827bd09bSSatish Balay { 994a501084fSBarry Smith int i, j, n, nel; 995a501084fSBarry Smith int *iptr_in, *iptr_out, *tree_elms, *elms; 996827bd09bSSatish Balay 9973fdc5746SBarry Smith PetscFunctionBegin; 998827bd09bSSatish Balay /* local work ptrs */ 999827bd09bSSatish Balay elms = gs->elms; 1000827bd09bSSatish Balay nel = gs->nel; 1001827bd09bSSatish Balay 1002827bd09bSSatish Balay /* how many via tree */ 1003827bd09bSSatish Balay gs->tree_nel = n = ntree; 1004827bd09bSSatish Balay gs->tree_elms = tree_elms = iptr_in = tree_buf; 1005a501084fSBarry Smith gs->tree_buf = (PetscScalar *) malloc(sizeof(PetscScalar)*n*vec_sz); 1006a501084fSBarry Smith gs->tree_work = (PetscScalar *) malloc(sizeof(PetscScalar)*n*vec_sz); 1007827bd09bSSatish Balay j=gs->tree_map_sz; 1008a501084fSBarry Smith gs->tree_map_in = iptr_in = (int*) malloc(sizeof(PetscInt)*(j+1)); 1009a501084fSBarry Smith gs->tree_map_out = iptr_out = (int*) malloc(sizeof(PetscInt)*(j+1)); 1010827bd09bSSatish Balay 1011827bd09bSSatish Balay /* search the longer of the two lists */ 1012827bd09bSSatish Balay /* note ... could save this info in get_ngh_buf and save searches */ 1013827bd09bSSatish Balay if (n<=nel) 1014827bd09bSSatish Balay { 1015827bd09bSSatish Balay /* bijective fct w/remap - search elm list */ 1016827bd09bSSatish Balay for (i=0; i<n; i++) 1017827bd09bSSatish Balay { 1018827bd09bSSatish Balay if ((j=ivec_binary_search(*tree_elms++,elms,nel))>=0) 1019827bd09bSSatish Balay {*iptr_in++ = j; *iptr_out++ = i;} 1020827bd09bSSatish Balay } 1021827bd09bSSatish Balay } 1022827bd09bSSatish Balay else 1023827bd09bSSatish Balay { 1024827bd09bSSatish Balay for (i=0; i<nel; i++) 1025827bd09bSSatish Balay { 1026827bd09bSSatish Balay if ((j=ivec_binary_search(*elms++,tree_elms,n))>=0) 1027827bd09bSSatish Balay {*iptr_in++ = i; *iptr_out++ = j;} 1028827bd09bSSatish Balay } 1029827bd09bSSatish Balay } 1030827bd09bSSatish Balay 1031827bd09bSSatish Balay /* sentinel */ 1032827bd09bSSatish Balay *iptr_in = *iptr_out = -1; 10333fdc5746SBarry Smith PetscFunctionReturn(0); 1034827bd09bSSatish Balay } 1035827bd09bSSatish Balay 1036827bd09bSSatish Balay 1037827bd09bSSatish Balay /****************************************************************************** 1038827bd09bSSatish Balay Function: gather_scatter 1039827bd09bSSatish Balay 1040827bd09bSSatish Balay Input : 1041827bd09bSSatish Balay Output: 1042827bd09bSSatish Balay Return: 1043827bd09bSSatish Balay Description: 1044827bd09bSSatish Balay ******************************************************************************/ 10450924e98cSBarry Smith static PetscErrorCode gs_gop_local_out( gs_id *gs, PetscScalar *vals) 1046827bd09bSSatish Balay { 1047a501084fSBarry Smith int *num, *map, **reduce; 1048a501084fSBarry Smith PetscScalar tmp; 1049827bd09bSSatish Balay 10503fdc5746SBarry Smith PetscFunctionBegin; 1051827bd09bSSatish Balay num = gs->num_gop_local_reduce; 1052827bd09bSSatish Balay reduce = gs->gop_local_reduce; 1053827bd09bSSatish Balay while ((map = *reduce++)) 1054827bd09bSSatish Balay { 1055827bd09bSSatish Balay /* wall */ 1056827bd09bSSatish Balay if (*num == 2) 1057827bd09bSSatish Balay { 1058827bd09bSSatish Balay num ++; 1059827bd09bSSatish Balay vals[map[1]] = vals[map[0]]; 1060827bd09bSSatish Balay } 1061827bd09bSSatish Balay /* corner shared by three elements */ 1062827bd09bSSatish Balay else if (*num == 3) 1063827bd09bSSatish Balay { 1064827bd09bSSatish Balay num ++; 1065827bd09bSSatish Balay vals[map[2]] = vals[map[1]] = vals[map[0]]; 1066827bd09bSSatish Balay } 1067827bd09bSSatish Balay /* corner shared by four elements */ 1068827bd09bSSatish Balay else if (*num == 4) 1069827bd09bSSatish Balay { 1070827bd09bSSatish Balay num ++; 1071827bd09bSSatish Balay vals[map[3]] = vals[map[2]] = vals[map[1]] = vals[map[0]]; 1072827bd09bSSatish Balay } 1073827bd09bSSatish Balay /* general case ... odd geoms ... 3D*/ 1074827bd09bSSatish Balay else 1075827bd09bSSatish Balay { 1076827bd09bSSatish Balay num++; 1077827bd09bSSatish Balay tmp = *(vals + *map++); 1078827bd09bSSatish Balay while (*map >= 0) 1079827bd09bSSatish Balay {*(vals + *map++) = tmp;} 1080827bd09bSSatish Balay } 1081827bd09bSSatish Balay } 10823fdc5746SBarry Smith PetscFunctionReturn(0); 1083827bd09bSSatish Balay } 1084827bd09bSSatish Balay 1085827bd09bSSatish Balay 1086827bd09bSSatish Balay 1087827bd09bSSatish Balay /****************************************************************************** 1088827bd09bSSatish Balay Function: gather_scatter 1089827bd09bSSatish Balay 1090827bd09bSSatish Balay Input : 1091827bd09bSSatish Balay Output: 1092827bd09bSSatish Balay Return: 1093827bd09bSSatish Balay Description: 1094827bd09bSSatish Balay ******************************************************************************/ 10950924e98cSBarry Smith PetscErrorCode gs_gop_binary(gs_ADT gs, PetscScalar *vals, rbfp fct) 1096827bd09bSSatish Balay { 10973fdc5746SBarry Smith PetscFunctionBegin; 1098827bd09bSSatish Balay /* local only operations!!! */ 1099827bd09bSSatish Balay if (gs->num_local) 1100827bd09bSSatish Balay {gs_gop_local_binary(gs,vals,fct);} 1101827bd09bSSatish Balay 1102827bd09bSSatish Balay /* if intersection tree/pairwise and local isn't empty */ 1103827bd09bSSatish Balay if (gs->num_local_gop) 1104827bd09bSSatish Balay { 1105827bd09bSSatish Balay gs_gop_local_in_binary(gs,vals,fct); 1106827bd09bSSatish Balay 1107827bd09bSSatish Balay /* pairwise */ 1108827bd09bSSatish Balay if (gs->num_pairs) 1109827bd09bSSatish Balay {gs_gop_pairwise_binary(gs,vals,fct);} 1110827bd09bSSatish Balay 1111827bd09bSSatish Balay /* tree */ 1112827bd09bSSatish Balay else if (gs->max_left_over) 1113827bd09bSSatish Balay {gs_gop_tree_binary(gs,vals,fct);} 1114827bd09bSSatish Balay 1115827bd09bSSatish Balay gs_gop_local_out(gs,vals); 1116827bd09bSSatish Balay } 1117827bd09bSSatish Balay /* if intersection tree/pairwise and local is empty */ 1118827bd09bSSatish Balay else 1119827bd09bSSatish Balay { 1120827bd09bSSatish Balay /* pairwise */ 1121827bd09bSSatish Balay if (gs->num_pairs) 1122827bd09bSSatish Balay {gs_gop_pairwise_binary(gs,vals,fct);} 1123827bd09bSSatish Balay 1124827bd09bSSatish Balay /* tree */ 1125827bd09bSSatish Balay else if (gs->max_left_over) 1126827bd09bSSatish Balay {gs_gop_tree_binary(gs,vals,fct);} 1127827bd09bSSatish Balay } 11283fdc5746SBarry Smith PetscFunctionReturn(0); 1129827bd09bSSatish Balay } 1130827bd09bSSatish Balay 1131827bd09bSSatish Balay 1132827bd09bSSatish Balay 1133827bd09bSSatish Balay /****************************************************************************** 1134827bd09bSSatish Balay Function: gather_scatter 1135827bd09bSSatish Balay 1136827bd09bSSatish Balay Input : 1137827bd09bSSatish Balay Output: 1138827bd09bSSatish Balay Return: 1139827bd09bSSatish Balay Description: 1140827bd09bSSatish Balay ******************************************************************************/ 11410924e98cSBarry Smith static PetscErrorCode gs_gop_local_binary( gs_id *gs, PetscScalar *vals, rbfp fct) 1142827bd09bSSatish Balay { 1143a501084fSBarry Smith int *num, *map, **reduce; 1144a501084fSBarry Smith PetscScalar tmp; 1145827bd09bSSatish Balay 11463fdc5746SBarry Smith PetscFunctionBegin; 1147827bd09bSSatish Balay num = gs->num_local_reduce; 1148827bd09bSSatish Balay reduce = gs->local_reduce; 1149827bd09bSSatish Balay while ((map = *reduce)) 1150827bd09bSSatish Balay { 1151827bd09bSSatish Balay num ++; 1152827bd09bSSatish Balay (*fct)(&tmp,NULL,1); 1153827bd09bSSatish Balay /* tmp = 0.0; */ 1154827bd09bSSatish Balay while (*map >= 0) 1155827bd09bSSatish Balay {(*fct)(&tmp,(vals + *map),1); map++;} 1156827bd09bSSatish Balay /* {tmp = (*fct)(tmp,*(vals + *map)); map++;} */ 1157827bd09bSSatish Balay 1158827bd09bSSatish Balay map = *reduce++; 1159827bd09bSSatish Balay while (*map >= 0) 1160827bd09bSSatish Balay {*(vals + *map++) = tmp;} 1161827bd09bSSatish Balay } 11623fdc5746SBarry Smith PetscFunctionReturn(0); 1163827bd09bSSatish Balay } 1164827bd09bSSatish Balay 1165827bd09bSSatish Balay 1166827bd09bSSatish Balay 1167827bd09bSSatish Balay /****************************************************************************** 1168827bd09bSSatish Balay Function: gather_scatter 1169827bd09bSSatish Balay 1170827bd09bSSatish Balay Input : 1171827bd09bSSatish Balay Output: 1172827bd09bSSatish Balay Return: 1173827bd09bSSatish Balay Description: 1174827bd09bSSatish Balay ******************************************************************************/ 11750924e98cSBarry Smith static PetscErrorCode gs_gop_local_in_binary( gs_id *gs, PetscScalar *vals, rbfp fct) 1176827bd09bSSatish Balay { 1177a501084fSBarry Smith int *num, *map, **reduce; 1178a501084fSBarry Smith PetscScalar *base; 1179827bd09bSSatish Balay 11803fdc5746SBarry Smith PetscFunctionBegin; 1181827bd09bSSatish Balay num = gs->num_gop_local_reduce; 1182827bd09bSSatish Balay 1183827bd09bSSatish Balay reduce = gs->gop_local_reduce; 1184827bd09bSSatish Balay while ((map = *reduce++)) 1185827bd09bSSatish Balay { 1186827bd09bSSatish Balay num++; 1187827bd09bSSatish Balay base = vals + *map++; 1188827bd09bSSatish Balay while (*map >= 0) 1189827bd09bSSatish Balay {(*fct)(base,(vals + *map),1); map++;} 1190827bd09bSSatish Balay } 11913fdc5746SBarry Smith PetscFunctionReturn(0); 1192827bd09bSSatish Balay } 1193827bd09bSSatish Balay 1194827bd09bSSatish Balay 1195827bd09bSSatish Balay 1196827bd09bSSatish Balay /****************************************************************************** 1197827bd09bSSatish Balay Function: gather_scatter 1198827bd09bSSatish Balay 1199827bd09bSSatish Balay VERSION 3 :: 1200827bd09bSSatish Balay 1201827bd09bSSatish Balay Input : 1202827bd09bSSatish Balay Output: 1203827bd09bSSatish Balay Return: 1204827bd09bSSatish Balay Description: 1205827bd09bSSatish Balay ******************************************************************************/ 12060924e98cSBarry Smith static PetscErrorCode gs_gop_pairwise_binary( gs_id *gs, PetscScalar *in_vals, 1207a501084fSBarry Smith rbfp fct) 1208827bd09bSSatish Balay { 1209a501084fSBarry Smith PetscScalar *dptr1, *dptr2, *dptr3, *in1, *in2; 1210a501084fSBarry Smith int *iptr, *msg_list, *msg_size, **msg_nodes; 1211a501084fSBarry Smith int *pw, *list, *size, **nodes; 1212827bd09bSSatish Balay MPI_Request *msg_ids_in, *msg_ids_out, *ids_in, *ids_out; 1213827bd09bSSatish Balay MPI_Status status; 12143fdc5746SBarry Smith PetscErrorCode ierr; 1215827bd09bSSatish Balay 12163fdc5746SBarry Smith PetscFunctionBegin; 1217a501084fSBarry Smith /* strip and load s */ 1218827bd09bSSatish Balay msg_list =list = gs->pair_list; 1219827bd09bSSatish Balay msg_size =size = gs->msg_sizes; 1220827bd09bSSatish Balay msg_nodes=nodes = gs->node_list; 1221827bd09bSSatish Balay iptr=pw = gs->pw_elm_list; 1222827bd09bSSatish Balay dptr1=dptr3 = gs->pw_vals; 1223827bd09bSSatish Balay msg_ids_in = ids_in = gs->msg_ids_in; 1224827bd09bSSatish Balay msg_ids_out = ids_out = gs->msg_ids_out; 1225827bd09bSSatish Balay dptr2 = gs->out; 1226827bd09bSSatish Balay in1=in2 = gs->in; 1227827bd09bSSatish Balay 1228827bd09bSSatish Balay /* post the receives */ 1229827bd09bSSatish Balay /* msg_nodes=nodes; */ 1230827bd09bSSatish Balay do 1231827bd09bSSatish Balay { 1232827bd09bSSatish Balay /* Should MPI_ANY_SOURCE be replaced by *list ? In that case do the 1233827bd09bSSatish Balay second one *list and do list++ afterwards */ 12343fdc5746SBarry Smith ierr = MPI_Irecv(in1, *size, MPIU_SCALAR, MPI_ANY_SOURCE, MSGTAG1 + *list++, gs->gs_comm, msg_ids_in++);CHKERRQ(ierr); 1235827bd09bSSatish Balay in1 += *size++; 1236827bd09bSSatish Balay } 1237827bd09bSSatish Balay while (*++msg_nodes); 1238827bd09bSSatish Balay msg_nodes=nodes; 1239827bd09bSSatish Balay 1240827bd09bSSatish Balay /* load gs values into in out gs buffers */ 1241827bd09bSSatish Balay while (*iptr >= 0) 1242827bd09bSSatish Balay {*dptr3++ = *(in_vals + *iptr++);} 1243827bd09bSSatish Balay 1244827bd09bSSatish Balay /* load out buffers and post the sends */ 1245827bd09bSSatish Balay while ((iptr = *msg_nodes++)) 1246827bd09bSSatish Balay { 1247827bd09bSSatish Balay dptr3 = dptr2; 1248827bd09bSSatish Balay while (*iptr >= 0) 1249827bd09bSSatish Balay {*dptr2++ = *(dptr1 + *iptr++);} 1250827bd09bSSatish Balay /* CHECK PERSISTENT COMMS MODE FOR ALL THIS STUFF */ 1251827bd09bSSatish Balay /* is msg_ids_out++ correct? */ 12523fdc5746SBarry Smith ierr = MPI_Isend(dptr3, *msg_size++, MPIU_SCALAR, *msg_list++, MSGTAG1+my_id, gs->gs_comm, msg_ids_out++);CHKERRQ(ierr); 1253827bd09bSSatish Balay } 1254827bd09bSSatish Balay 1255827bd09bSSatish Balay if (gs->max_left_over) 1256827bd09bSSatish Balay {gs_gop_tree_binary(gs,in_vals,fct);} 1257827bd09bSSatish Balay 1258827bd09bSSatish Balay /* process the received data */ 1259827bd09bSSatish Balay msg_nodes=nodes; 1260827bd09bSSatish Balay while ((iptr = *nodes++)) 1261827bd09bSSatish Balay { 1262827bd09bSSatish Balay /* Should I check the return value of MPI_Wait() or status? */ 1263827bd09bSSatish Balay /* Can this loop be replaced by a call to MPI_Waitall()? */ 12643fdc5746SBarry Smith ierr = MPI_Wait(ids_in++, &status);CHKERRQ(ierr); 1265827bd09bSSatish Balay while (*iptr >= 0) 1266827bd09bSSatish Balay {(*fct)((dptr1 + *iptr),in2,1); iptr++; in2++;} 1267827bd09bSSatish Balay /* {*(dptr1 + *iptr) = (*fct)(*(dptr1 + *iptr),*in2); iptr++; in2++;} */ 1268827bd09bSSatish Balay } 1269827bd09bSSatish Balay 1270827bd09bSSatish Balay /* replace vals */ 1271827bd09bSSatish Balay while (*pw >= 0) 1272827bd09bSSatish Balay {*(in_vals + *pw++) = *dptr1++;} 1273827bd09bSSatish Balay 1274827bd09bSSatish Balay /* clear isend message handles */ 1275827bd09bSSatish Balay /* This changed for clarity though it could be the same */ 1276827bd09bSSatish Balay while (*msg_nodes++) 1277827bd09bSSatish Balay /* Should I check the return value of MPI_Wait() or status? */ 1278827bd09bSSatish Balay /* Can this loop be replaced by a call to MPI_Waitall()? */ 12793fdc5746SBarry Smith {ierr = MPI_Wait(ids_out++, &status);CHKERRQ(ierr);} 12803fdc5746SBarry Smith PetscFunctionReturn(0); 1281827bd09bSSatish Balay } 1282827bd09bSSatish Balay 1283827bd09bSSatish Balay 1284827bd09bSSatish Balay 1285827bd09bSSatish Balay /****************************************************************************** 1286827bd09bSSatish Balay Function: gather_scatter 1287827bd09bSSatish Balay 1288827bd09bSSatish Balay Input : 1289827bd09bSSatish Balay Output: 1290827bd09bSSatish Balay Return: 1291827bd09bSSatish Balay Description: 1292827bd09bSSatish Balay ******************************************************************************/ 12930924e98cSBarry Smith static PetscErrorCode gs_gop_tree_binary(gs_id *gs, PetscScalar *vals, rbfp fct) 1294827bd09bSSatish Balay { 1295827bd09bSSatish Balay int size; 1296827bd09bSSatish Balay int *in, *out; 1297a501084fSBarry Smith PetscScalar *buf, *work; 1298827bd09bSSatish Balay 12993fdc5746SBarry Smith PetscFunctionBegin; 1300827bd09bSSatish Balay in = gs->tree_map_in; 1301827bd09bSSatish Balay out = gs->tree_map_out; 1302827bd09bSSatish Balay buf = gs->tree_buf; 1303827bd09bSSatish Balay work = gs->tree_work; 1304827bd09bSSatish Balay size = gs->tree_nel; 1305827bd09bSSatish Balay 1306827bd09bSSatish Balay /* load vals vector w/identity */ 1307827bd09bSSatish Balay (*fct)(buf,NULL,size); 1308827bd09bSSatish Balay 1309827bd09bSSatish Balay /* load my contribution into val vector */ 1310827bd09bSSatish Balay while (*in >= 0) 1311827bd09bSSatish Balay {(*fct)((buf + *out++),(vals + *in++),-1);} 1312827bd09bSSatish Balay 1313a501084fSBarry Smith gfop(buf,work,size,(vbfp)fct,MPIU_SCALAR,0); 1314827bd09bSSatish Balay 1315827bd09bSSatish Balay in = gs->tree_map_in; 1316827bd09bSSatish Balay out = gs->tree_map_out; 1317827bd09bSSatish Balay while (*in >= 0) 1318827bd09bSSatish Balay {(*fct)((vals + *in++),(buf + *out++),-1);} 13193fdc5746SBarry Smith PetscFunctionReturn(0); 1320827bd09bSSatish Balay } 1321827bd09bSSatish Balay 1322827bd09bSSatish Balay 1323827bd09bSSatish Balay 1324827bd09bSSatish Balay 1325827bd09bSSatish Balay /****************************************************************************** 1326827bd09bSSatish Balay Function: gather_scatter 1327827bd09bSSatish Balay 1328827bd09bSSatish Balay Input : 1329827bd09bSSatish Balay Output: 1330827bd09bSSatish Balay Return: 1331827bd09bSSatish Balay Description: 1332827bd09bSSatish Balay ******************************************************************************/ 13330924e98cSBarry Smith PetscErrorCode gs_gop( gs_id *gs, PetscScalar *vals, const char *op) 1334827bd09bSSatish Balay { 13353fdc5746SBarry Smith PetscFunctionBegin; 1336827bd09bSSatish Balay switch (*op) { 1337827bd09bSSatish Balay case '+': 1338827bd09bSSatish Balay gs_gop_plus(gs,vals); 1339827bd09bSSatish Balay break; 1340827bd09bSSatish Balay case '*': 1341827bd09bSSatish Balay gs_gop_times(gs,vals); 1342827bd09bSSatish Balay break; 1343827bd09bSSatish Balay case 'a': 1344827bd09bSSatish Balay gs_gop_min_abs(gs,vals); 1345827bd09bSSatish Balay break; 1346827bd09bSSatish Balay case 'A': 1347827bd09bSSatish Balay gs_gop_max_abs(gs,vals); 1348827bd09bSSatish Balay break; 1349827bd09bSSatish Balay case 'e': 1350827bd09bSSatish Balay gs_gop_exists(gs,vals); 1351827bd09bSSatish Balay break; 1352827bd09bSSatish Balay case 'm': 1353827bd09bSSatish Balay gs_gop_min(gs,vals); 1354827bd09bSSatish Balay break; 1355827bd09bSSatish Balay case 'M': 1356827bd09bSSatish Balay gs_gop_max(gs,vals); break; 1357827bd09bSSatish Balay /* 1358827bd09bSSatish Balay if (*(op+1)=='\0') 1359827bd09bSSatish Balay {gs_gop_max(gs,vals); break;} 1360827bd09bSSatish Balay else if (*(op+1)=='X') 1361827bd09bSSatish Balay {gs_gop_max_abs(gs,vals); break;} 1362827bd09bSSatish Balay else if (*(op+1)=='N') 1363827bd09bSSatish Balay {gs_gop_min_abs(gs,vals); break;} 1364827bd09bSSatish Balay */ 1365827bd09bSSatish Balay default: 1366827bd09bSSatish Balay error_msg_warning("gs_gop() :: %c is not a valid op",op[0]); 1367827bd09bSSatish Balay error_msg_warning("gs_gop() :: default :: plus"); 1368827bd09bSSatish Balay gs_gop_plus(gs,vals); 1369827bd09bSSatish Balay break; 1370827bd09bSSatish Balay } 13713fdc5746SBarry Smith PetscFunctionReturn(0); 1372827bd09bSSatish Balay } 1373827bd09bSSatish Balay 1374827bd09bSSatish Balay 1375827bd09bSSatish Balay /****************************************************************************** 1376827bd09bSSatish Balay Function: gather_scatter 1377827bd09bSSatish Balay 1378827bd09bSSatish Balay Input : 1379827bd09bSSatish Balay Output: 1380827bd09bSSatish Balay Return: 1381827bd09bSSatish Balay Description: 1382827bd09bSSatish Balay ******************************************************************************/ 13830924e98cSBarry Smith static PetscErrorCode gs_gop_exists( gs_id *gs, PetscScalar *vals) 1384827bd09bSSatish Balay { 13853fdc5746SBarry Smith PetscFunctionBegin; 1386827bd09bSSatish Balay /* local only operations!!! */ 1387827bd09bSSatish Balay if (gs->num_local) 1388827bd09bSSatish Balay {gs_gop_local_exists(gs,vals);} 1389827bd09bSSatish Balay 1390827bd09bSSatish Balay /* if intersection tree/pairwise and local isn't empty */ 1391827bd09bSSatish Balay if (gs->num_local_gop) 1392827bd09bSSatish Balay { 1393827bd09bSSatish Balay gs_gop_local_in_exists(gs,vals); 1394827bd09bSSatish Balay 1395827bd09bSSatish Balay /* pairwise */ 1396827bd09bSSatish Balay if (gs->num_pairs) 1397827bd09bSSatish Balay {gs_gop_pairwise_exists(gs,vals);} 1398827bd09bSSatish Balay 1399827bd09bSSatish Balay /* tree */ 1400827bd09bSSatish Balay else if (gs->max_left_over) 1401827bd09bSSatish Balay {gs_gop_tree_exists(gs,vals);} 1402827bd09bSSatish Balay 1403827bd09bSSatish Balay gs_gop_local_out(gs,vals); 1404827bd09bSSatish Balay } 1405827bd09bSSatish Balay /* if intersection tree/pairwise and local is empty */ 1406827bd09bSSatish Balay else 1407827bd09bSSatish Balay { 1408827bd09bSSatish Balay /* pairwise */ 1409827bd09bSSatish Balay if (gs->num_pairs) 1410827bd09bSSatish Balay {gs_gop_pairwise_exists(gs,vals);} 1411827bd09bSSatish Balay 1412827bd09bSSatish Balay /* tree */ 1413827bd09bSSatish Balay else if (gs->max_left_over) 1414827bd09bSSatish Balay {gs_gop_tree_exists(gs,vals);} 1415827bd09bSSatish Balay } 14163fdc5746SBarry Smith PetscFunctionReturn(0); 1417827bd09bSSatish Balay } 1418827bd09bSSatish Balay 1419827bd09bSSatish Balay 1420827bd09bSSatish Balay 1421827bd09bSSatish Balay /****************************************************************************** 1422827bd09bSSatish Balay Function: gather_scatter 1423827bd09bSSatish Balay 1424827bd09bSSatish Balay Input : 1425827bd09bSSatish Balay Output: 1426827bd09bSSatish Balay Return: 1427827bd09bSSatish Balay Description: 1428827bd09bSSatish Balay ******************************************************************************/ 14290924e98cSBarry Smith static PetscErrorCode gs_gop_local_exists( gs_id *gs, PetscScalar *vals) 1430827bd09bSSatish Balay { 1431a501084fSBarry Smith int *num, *map, **reduce; 1432a501084fSBarry Smith PetscScalar tmp; 1433827bd09bSSatish Balay 14343fdc5746SBarry Smith PetscFunctionBegin; 1435827bd09bSSatish Balay num = gs->num_local_reduce; 1436827bd09bSSatish Balay reduce = gs->local_reduce; 1437827bd09bSSatish Balay while ((map = *reduce)) 1438827bd09bSSatish Balay { 1439827bd09bSSatish Balay num ++; 1440827bd09bSSatish Balay tmp = 0.0; 1441827bd09bSSatish Balay while (*map >= 0) 1442827bd09bSSatish Balay {tmp = EXISTS(tmp,*(vals + *map)); map++;} 1443827bd09bSSatish Balay 1444827bd09bSSatish Balay map = *reduce++; 1445827bd09bSSatish Balay while (*map >= 0) 1446827bd09bSSatish Balay {*(vals + *map++) = tmp;} 1447827bd09bSSatish Balay } 14483fdc5746SBarry Smith PetscFunctionReturn(0); 1449827bd09bSSatish Balay } 1450827bd09bSSatish Balay 1451827bd09bSSatish Balay 1452827bd09bSSatish Balay 1453827bd09bSSatish Balay /****************************************************************************** 1454827bd09bSSatish Balay Function: gather_scatter 1455827bd09bSSatish Balay 1456827bd09bSSatish Balay Input : 1457827bd09bSSatish Balay Output: 1458827bd09bSSatish Balay Return: 1459827bd09bSSatish Balay Description: 1460827bd09bSSatish Balay ******************************************************************************/ 14610924e98cSBarry Smith static PetscErrorCode gs_gop_local_in_exists( gs_id *gs, PetscScalar *vals) 1462827bd09bSSatish Balay { 1463a501084fSBarry Smith int *num, *map, **reduce; 1464a501084fSBarry Smith PetscScalar *base; 1465827bd09bSSatish Balay 14663fdc5746SBarry Smith PetscFunctionBegin; 1467827bd09bSSatish Balay num = gs->num_gop_local_reduce; 1468827bd09bSSatish Balay reduce = gs->gop_local_reduce; 1469827bd09bSSatish Balay while ((map = *reduce++)) 1470827bd09bSSatish Balay { 1471827bd09bSSatish Balay num++; 1472827bd09bSSatish Balay base = vals + *map++; 1473827bd09bSSatish Balay while (*map >= 0) 1474827bd09bSSatish Balay {*base = EXISTS(*base,*(vals + *map)); map++;} 1475827bd09bSSatish Balay } 14763fdc5746SBarry Smith PetscFunctionReturn(0); 1477827bd09bSSatish Balay } 1478827bd09bSSatish Balay 1479827bd09bSSatish Balay 1480827bd09bSSatish Balay 1481827bd09bSSatish Balay /****************************************************************************** 1482827bd09bSSatish Balay Function: gather_scatter 1483827bd09bSSatish Balay 1484827bd09bSSatish Balay VERSION 3 :: 1485827bd09bSSatish Balay 1486827bd09bSSatish Balay Input : 1487827bd09bSSatish Balay Output: 1488827bd09bSSatish Balay Return: 1489827bd09bSSatish Balay Description: 1490827bd09bSSatish Balay ******************************************************************************/ 14910924e98cSBarry Smith static PetscErrorCode gs_gop_pairwise_exists( gs_id *gs, PetscScalar *in_vals) 1492827bd09bSSatish Balay { 1493a501084fSBarry Smith PetscScalar *dptr1, *dptr2, *dptr3, *in1, *in2; 1494a501084fSBarry Smith int *iptr, *msg_list, *msg_size, **msg_nodes; 1495a501084fSBarry Smith int *pw, *list, *size, **nodes; 1496827bd09bSSatish Balay MPI_Request *msg_ids_in, *msg_ids_out, *ids_in, *ids_out; 1497827bd09bSSatish Balay MPI_Status status; 14983fdc5746SBarry Smith PetscErrorCode ierr; 1499827bd09bSSatish Balay 15003fdc5746SBarry Smith PetscFunctionBegin; 1501a501084fSBarry Smith /* strip and load s */ 1502827bd09bSSatish Balay msg_list =list = gs->pair_list; 1503827bd09bSSatish Balay msg_size =size = gs->msg_sizes; 1504827bd09bSSatish Balay msg_nodes=nodes = gs->node_list; 1505827bd09bSSatish Balay iptr=pw = gs->pw_elm_list; 1506827bd09bSSatish Balay dptr1=dptr3 = gs->pw_vals; 1507827bd09bSSatish Balay msg_ids_in = ids_in = gs->msg_ids_in; 1508827bd09bSSatish Balay msg_ids_out = ids_out = gs->msg_ids_out; 1509827bd09bSSatish Balay dptr2 = gs->out; 1510827bd09bSSatish Balay in1=in2 = gs->in; 1511827bd09bSSatish Balay 1512827bd09bSSatish Balay /* post the receives */ 1513827bd09bSSatish Balay /* msg_nodes=nodes; */ 1514827bd09bSSatish Balay do 1515827bd09bSSatish Balay { 1516827bd09bSSatish Balay /* Should MPI_ANY_SOURCE be replaced by *list ? In that case do the 1517827bd09bSSatish Balay second one *list and do list++ afterwards */ 15183fdc5746SBarry Smith ierr = MPI_Irecv(in1, *size, MPIU_SCALAR, MPI_ANY_SOURCE, MSGTAG1 + *list++, gs->gs_comm, msg_ids_in++);CHKERRQ(ierr); 1519827bd09bSSatish Balay in1 += *size++; 1520827bd09bSSatish Balay } 1521827bd09bSSatish Balay while (*++msg_nodes); 1522827bd09bSSatish Balay msg_nodes=nodes; 1523827bd09bSSatish Balay 1524827bd09bSSatish Balay /* load gs values into in out gs buffers */ 1525827bd09bSSatish Balay while (*iptr >= 0) 1526827bd09bSSatish Balay {*dptr3++ = *(in_vals + *iptr++);} 1527827bd09bSSatish Balay 1528827bd09bSSatish Balay /* load out buffers and post the sends */ 1529827bd09bSSatish Balay while ((iptr = *msg_nodes++)) 1530827bd09bSSatish Balay { 1531827bd09bSSatish Balay dptr3 = dptr2; 1532827bd09bSSatish Balay while (*iptr >= 0) 1533827bd09bSSatish Balay {*dptr2++ = *(dptr1 + *iptr++);} 1534827bd09bSSatish Balay /* CHECK PERSISTENT COMMS MODE FOR ALL THIS STUFF */ 1535827bd09bSSatish Balay /* is msg_ids_out++ correct? */ 15363fdc5746SBarry Smith ierr = MPI_Isend(dptr3, *msg_size++, MPIU_SCALAR, *msg_list++, MSGTAG1+my_id, gs->gs_comm, msg_ids_out++);CHKERRQ(ierr); 1537827bd09bSSatish Balay } 1538827bd09bSSatish Balay 1539827bd09bSSatish Balay if (gs->max_left_over) 1540827bd09bSSatish Balay {gs_gop_tree_exists(gs,in_vals);} 1541827bd09bSSatish Balay 1542827bd09bSSatish Balay /* process the received data */ 1543827bd09bSSatish Balay msg_nodes=nodes; 1544827bd09bSSatish Balay while ((iptr = *nodes++)) 1545827bd09bSSatish Balay { 1546827bd09bSSatish Balay /* Should I check the return value of MPI_Wait() or status? */ 1547827bd09bSSatish Balay /* Can this loop be replaced by a call to MPI_Waitall()? */ 15483fdc5746SBarry Smith ierr = MPI_Wait(ids_in++, &status);CHKERRQ(ierr); 1549827bd09bSSatish Balay while (*iptr >= 0) 1550827bd09bSSatish Balay {*(dptr1 + *iptr) = EXISTS(*(dptr1 + *iptr),*in2); iptr++; in2++;} 1551827bd09bSSatish Balay } 1552827bd09bSSatish Balay 1553827bd09bSSatish Balay /* replace vals */ 1554827bd09bSSatish Balay while (*pw >= 0) 1555827bd09bSSatish Balay {*(in_vals + *pw++) = *dptr1++;} 1556827bd09bSSatish Balay 1557827bd09bSSatish Balay /* clear isend message handles */ 1558827bd09bSSatish Balay /* This changed for clarity though it could be the same */ 1559827bd09bSSatish Balay while (*msg_nodes++) 1560827bd09bSSatish Balay /* Should I check the return value of MPI_Wait() or status? */ 1561827bd09bSSatish Balay /* Can this loop be replaced by a call to MPI_Waitall()? */ 15623fdc5746SBarry Smith {ierr = MPI_Wait(ids_out++, &status);CHKERRQ(ierr);} 15633fdc5746SBarry Smith PetscFunctionReturn(0); 1564827bd09bSSatish Balay } 1565827bd09bSSatish Balay 1566827bd09bSSatish Balay 1567827bd09bSSatish Balay 1568827bd09bSSatish Balay /****************************************************************************** 1569827bd09bSSatish Balay Function: gather_scatter 1570827bd09bSSatish Balay 1571827bd09bSSatish Balay Input : 1572827bd09bSSatish Balay Output: 1573827bd09bSSatish Balay Return: 1574827bd09bSSatish Balay Description: 1575827bd09bSSatish Balay ******************************************************************************/ 15760924e98cSBarry Smith static PetscErrorCode gs_gop_tree_exists(gs_id *gs, PetscScalar *vals) 1577827bd09bSSatish Balay { 1578827bd09bSSatish Balay int size; 1579827bd09bSSatish Balay int *in, *out; 1580a501084fSBarry Smith PetscScalar *buf, *work; 1581827bd09bSSatish Balay int op[] = {GL_EXISTS,0}; 1582827bd09bSSatish Balay 15833fdc5746SBarry Smith PetscFunctionBegin; 1584827bd09bSSatish Balay in = gs->tree_map_in; 1585827bd09bSSatish Balay out = gs->tree_map_out; 1586827bd09bSSatish Balay buf = gs->tree_buf; 1587827bd09bSSatish Balay work = gs->tree_work; 1588827bd09bSSatish Balay size = gs->tree_nel; 1589827bd09bSSatish Balay 1590827bd09bSSatish Balay rvec_zero(buf,size); 1591827bd09bSSatish Balay 1592827bd09bSSatish Balay while (*in >= 0) 1593827bd09bSSatish Balay { 1594827bd09bSSatish Balay /* 1595827bd09bSSatish Balay printf("%d :: out=%d\n",my_id,*out); 1596827bd09bSSatish Balay printf("%d :: in=%d\n",my_id,*in); 1597827bd09bSSatish Balay */ 1598827bd09bSSatish Balay *(buf + *out++) = *(vals + *in++); 1599827bd09bSSatish Balay } 1600827bd09bSSatish Balay 1601827bd09bSSatish Balay grop(buf,work,size,op); 1602827bd09bSSatish Balay 1603827bd09bSSatish Balay in = gs->tree_map_in; 1604827bd09bSSatish Balay out = gs->tree_map_out; 1605827bd09bSSatish Balay 1606827bd09bSSatish Balay while (*in >= 0) 1607827bd09bSSatish Balay {*(vals + *in++) = *(buf + *out++);} 16083fdc5746SBarry Smith PetscFunctionReturn(0); 1609827bd09bSSatish Balay } 1610827bd09bSSatish Balay 1611827bd09bSSatish Balay 1612827bd09bSSatish Balay 1613827bd09bSSatish Balay /****************************************************************************** 1614827bd09bSSatish Balay Function: gather_scatter 1615827bd09bSSatish Balay 1616827bd09bSSatish Balay Input : 1617827bd09bSSatish Balay Output: 1618827bd09bSSatish Balay Return: 1619827bd09bSSatish Balay Description: 1620827bd09bSSatish Balay ******************************************************************************/ 16210924e98cSBarry Smith static PetscErrorCode gs_gop_max_abs( gs_id *gs, PetscScalar *vals) 1622827bd09bSSatish Balay { 16233fdc5746SBarry Smith PetscFunctionBegin; 1624827bd09bSSatish Balay /* local only operations!!! */ 1625827bd09bSSatish Balay if (gs->num_local) 1626827bd09bSSatish Balay {gs_gop_local_max_abs(gs,vals);} 1627827bd09bSSatish Balay 1628827bd09bSSatish Balay /* if intersection tree/pairwise and local isn't empty */ 1629827bd09bSSatish Balay if (gs->num_local_gop) 1630827bd09bSSatish Balay { 1631827bd09bSSatish Balay gs_gop_local_in_max_abs(gs,vals); 1632827bd09bSSatish Balay 1633827bd09bSSatish Balay /* pairwise */ 1634827bd09bSSatish Balay if (gs->num_pairs) 1635827bd09bSSatish Balay {gs_gop_pairwise_max_abs(gs,vals);} 1636827bd09bSSatish Balay 1637827bd09bSSatish Balay /* tree */ 1638827bd09bSSatish Balay else if (gs->max_left_over) 1639827bd09bSSatish Balay {gs_gop_tree_max_abs(gs,vals);} 1640827bd09bSSatish Balay 1641827bd09bSSatish Balay gs_gop_local_out(gs,vals); 1642827bd09bSSatish Balay } 1643827bd09bSSatish Balay /* if intersection tree/pairwise and local is empty */ 1644827bd09bSSatish Balay else 1645827bd09bSSatish Balay { 1646827bd09bSSatish Balay /* pairwise */ 1647827bd09bSSatish Balay if (gs->num_pairs) 1648827bd09bSSatish Balay {gs_gop_pairwise_max_abs(gs,vals);} 1649827bd09bSSatish Balay 1650827bd09bSSatish Balay /* tree */ 1651827bd09bSSatish Balay else if (gs->max_left_over) 1652827bd09bSSatish Balay {gs_gop_tree_max_abs(gs,vals);} 1653827bd09bSSatish Balay } 16543fdc5746SBarry Smith PetscFunctionReturn(0); 1655827bd09bSSatish Balay } 1656827bd09bSSatish Balay 1657827bd09bSSatish Balay 1658827bd09bSSatish Balay 1659827bd09bSSatish Balay /****************************************************************************** 1660827bd09bSSatish Balay Function: gather_scatter 1661827bd09bSSatish Balay 1662827bd09bSSatish Balay Input : 1663827bd09bSSatish Balay Output: 1664827bd09bSSatish Balay Return: 1665827bd09bSSatish Balay Description: 1666827bd09bSSatish Balay ******************************************************************************/ 16670924e98cSBarry Smith static PetscErrorCode gs_gop_local_max_abs( gs_id *gs, PetscScalar *vals) 1668827bd09bSSatish Balay { 1669a501084fSBarry Smith int *num, *map, **reduce; 1670a501084fSBarry Smith PetscScalar tmp; 1671827bd09bSSatish Balay 16723fdc5746SBarry Smith PetscFunctionBegin; 1673827bd09bSSatish Balay num = gs->num_local_reduce; 1674827bd09bSSatish Balay reduce = gs->local_reduce; 1675827bd09bSSatish Balay while ((map = *reduce)) 1676827bd09bSSatish Balay { 1677827bd09bSSatish Balay num ++; 1678827bd09bSSatish Balay tmp = 0.0; 1679827bd09bSSatish Balay while (*map >= 0) 1680827bd09bSSatish Balay {tmp = MAX_FABS(tmp,*(vals + *map)); map++;} 1681827bd09bSSatish Balay 1682827bd09bSSatish Balay map = *reduce++; 1683827bd09bSSatish Balay while (*map >= 0) 1684827bd09bSSatish Balay {*(vals + *map++) = tmp;} 1685827bd09bSSatish Balay } 16863fdc5746SBarry Smith PetscFunctionReturn(0); 1687827bd09bSSatish Balay } 1688827bd09bSSatish Balay 1689827bd09bSSatish Balay 1690827bd09bSSatish Balay 1691827bd09bSSatish Balay /****************************************************************************** 1692827bd09bSSatish Balay Function: gather_scatter 1693827bd09bSSatish Balay 1694827bd09bSSatish Balay Input : 1695827bd09bSSatish Balay Output: 1696827bd09bSSatish Balay Return: 1697827bd09bSSatish Balay Description: 1698827bd09bSSatish Balay ******************************************************************************/ 16990924e98cSBarry Smith static PetscErrorCode gs_gop_local_in_max_abs( gs_id *gs, PetscScalar *vals) 1700827bd09bSSatish Balay { 1701a501084fSBarry Smith int *num, *map, **reduce; 1702a501084fSBarry Smith PetscScalar *base; 1703827bd09bSSatish Balay 17043fdc5746SBarry Smith PetscFunctionBegin; 1705827bd09bSSatish Balay num = gs->num_gop_local_reduce; 1706827bd09bSSatish Balay reduce = gs->gop_local_reduce; 1707827bd09bSSatish Balay while ((map = *reduce++)) 1708827bd09bSSatish Balay { 1709827bd09bSSatish Balay num++; 1710827bd09bSSatish Balay base = vals + *map++; 1711827bd09bSSatish Balay while (*map >= 0) 1712827bd09bSSatish Balay {*base = MAX_FABS(*base,*(vals + *map)); map++;} 1713827bd09bSSatish Balay } 17143fdc5746SBarry Smith PetscFunctionReturn(0); 1715827bd09bSSatish Balay } 1716827bd09bSSatish Balay 1717827bd09bSSatish Balay 1718827bd09bSSatish Balay 1719827bd09bSSatish Balay /****************************************************************************** 1720827bd09bSSatish Balay Function: gather_scatter 1721827bd09bSSatish Balay 1722827bd09bSSatish Balay VERSION 3 :: 1723827bd09bSSatish Balay 1724827bd09bSSatish Balay Input : 1725827bd09bSSatish Balay Output: 1726827bd09bSSatish Balay Return: 1727827bd09bSSatish Balay Description: 1728827bd09bSSatish Balay ******************************************************************************/ 17290924e98cSBarry Smith static PetscErrorCode gs_gop_pairwise_max_abs( gs_id *gs, PetscScalar *in_vals) 1730827bd09bSSatish Balay { 1731a501084fSBarry Smith PetscScalar *dptr1, *dptr2, *dptr3, *in1, *in2; 1732a501084fSBarry Smith int *iptr, *msg_list, *msg_size, **msg_nodes; 1733a501084fSBarry Smith int *pw, *list, *size, **nodes; 1734827bd09bSSatish Balay MPI_Request *msg_ids_in, *msg_ids_out, *ids_in, *ids_out; 1735827bd09bSSatish Balay MPI_Status status; 17363fdc5746SBarry Smith PetscErrorCode ierr; 1737827bd09bSSatish Balay 17383fdc5746SBarry Smith PetscFunctionBegin; 1739a501084fSBarry Smith /* strip and load s */ 1740827bd09bSSatish Balay msg_list =list = gs->pair_list; 1741827bd09bSSatish Balay msg_size =size = gs->msg_sizes; 1742827bd09bSSatish Balay msg_nodes=nodes = gs->node_list; 1743827bd09bSSatish Balay iptr=pw = gs->pw_elm_list; 1744827bd09bSSatish Balay dptr1=dptr3 = gs->pw_vals; 1745827bd09bSSatish Balay msg_ids_in = ids_in = gs->msg_ids_in; 1746827bd09bSSatish Balay msg_ids_out = ids_out = gs->msg_ids_out; 1747827bd09bSSatish Balay dptr2 = gs->out; 1748827bd09bSSatish Balay in1=in2 = gs->in; 1749827bd09bSSatish Balay 1750827bd09bSSatish Balay /* post the receives */ 1751827bd09bSSatish Balay /* msg_nodes=nodes; */ 1752827bd09bSSatish Balay do 1753827bd09bSSatish Balay { 1754827bd09bSSatish Balay /* Should MPI_ANY_SOURCE be replaced by *list ? In that case do the 1755827bd09bSSatish Balay second one *list and do list++ afterwards */ 17563fdc5746SBarry Smith ierr = MPI_Irecv(in1, *size, MPIU_SCALAR, MPI_ANY_SOURCE, MSGTAG1 + *list++, gs->gs_comm, msg_ids_in++);CHKERRQ(ierr); 1757827bd09bSSatish Balay in1 += *size++; 1758827bd09bSSatish Balay } 1759827bd09bSSatish Balay while (*++msg_nodes); 1760827bd09bSSatish Balay msg_nodes=nodes; 1761827bd09bSSatish Balay 1762827bd09bSSatish Balay /* load gs values into in out gs buffers */ 1763827bd09bSSatish Balay while (*iptr >= 0) 1764827bd09bSSatish Balay {*dptr3++ = *(in_vals + *iptr++);} 1765827bd09bSSatish Balay 1766827bd09bSSatish Balay /* load out buffers and post the sends */ 1767827bd09bSSatish Balay while ((iptr = *msg_nodes++)) 1768827bd09bSSatish Balay { 1769827bd09bSSatish Balay dptr3 = dptr2; 1770827bd09bSSatish Balay while (*iptr >= 0) 1771827bd09bSSatish Balay {*dptr2++ = *(dptr1 + *iptr++);} 1772827bd09bSSatish Balay /* CHECK PERSISTENT COMMS MODE FOR ALL THIS STUFF */ 1773827bd09bSSatish Balay /* is msg_ids_out++ correct? */ 17743fdc5746SBarry Smith ierr = MPI_Isend(dptr3, *msg_size++, MPIU_SCALAR, *msg_list++, MSGTAG1+my_id, gs->gs_comm, msg_ids_out++);CHKERRQ(ierr); 1775827bd09bSSatish Balay } 1776827bd09bSSatish Balay 1777827bd09bSSatish Balay if (gs->max_left_over) 1778827bd09bSSatish Balay {gs_gop_tree_max_abs(gs,in_vals);} 1779827bd09bSSatish Balay 1780827bd09bSSatish Balay /* process the received data */ 1781827bd09bSSatish Balay msg_nodes=nodes; 1782827bd09bSSatish Balay while ((iptr = *nodes++)) 1783827bd09bSSatish Balay { 1784827bd09bSSatish Balay /* Should I check the return value of MPI_Wait() or status? */ 1785827bd09bSSatish Balay /* Can this loop be replaced by a call to MPI_Waitall()? */ 17863fdc5746SBarry Smith ierr = MPI_Wait(ids_in++, &status);CHKERRQ(ierr); 1787827bd09bSSatish Balay while (*iptr >= 0) 1788827bd09bSSatish Balay {*(dptr1 + *iptr) = MAX_FABS(*(dptr1 + *iptr),*in2); iptr++; in2++;} 1789827bd09bSSatish Balay } 1790827bd09bSSatish Balay 1791827bd09bSSatish Balay /* replace vals */ 1792827bd09bSSatish Balay while (*pw >= 0) 1793827bd09bSSatish Balay {*(in_vals + *pw++) = *dptr1++;} 1794827bd09bSSatish Balay 1795827bd09bSSatish Balay /* clear isend message handles */ 1796827bd09bSSatish Balay /* This changed for clarity though it could be the same */ 1797827bd09bSSatish Balay while (*msg_nodes++) 1798827bd09bSSatish Balay /* Should I check the return value of MPI_Wait() or status? */ 1799827bd09bSSatish Balay /* Can this loop be replaced by a call to MPI_Waitall()? */ 18003fdc5746SBarry Smith {ierr = MPI_Wait(ids_out++, &status);CHKERRQ(ierr);} 18013fdc5746SBarry Smith PetscFunctionReturn(0); 1802827bd09bSSatish Balay } 1803827bd09bSSatish Balay 1804827bd09bSSatish Balay 1805827bd09bSSatish Balay 1806827bd09bSSatish Balay /****************************************************************************** 1807827bd09bSSatish Balay Function: gather_scatter 1808827bd09bSSatish Balay 1809827bd09bSSatish Balay Input : 1810827bd09bSSatish Balay Output: 1811827bd09bSSatish Balay Return: 1812827bd09bSSatish Balay Description: 1813827bd09bSSatish Balay ******************************************************************************/ 18140924e98cSBarry Smith static PetscErrorCode gs_gop_tree_max_abs(gs_id *gs, PetscScalar *vals) 1815827bd09bSSatish Balay { 1816827bd09bSSatish Balay int size; 1817827bd09bSSatish Balay int *in, *out; 1818a501084fSBarry Smith PetscScalar *buf, *work; 1819827bd09bSSatish Balay int op[] = {GL_MAX_ABS,0}; 1820827bd09bSSatish Balay 18213fdc5746SBarry Smith PetscFunctionBegin; 1822827bd09bSSatish Balay in = gs->tree_map_in; 1823827bd09bSSatish Balay out = gs->tree_map_out; 1824827bd09bSSatish Balay buf = gs->tree_buf; 1825827bd09bSSatish Balay work = gs->tree_work; 1826827bd09bSSatish Balay size = gs->tree_nel; 1827827bd09bSSatish Balay 1828827bd09bSSatish Balay rvec_zero(buf,size); 1829827bd09bSSatish Balay 1830827bd09bSSatish Balay while (*in >= 0) 1831827bd09bSSatish Balay { 1832827bd09bSSatish Balay /* 1833827bd09bSSatish Balay printf("%d :: out=%d\n",my_id,*out); 1834827bd09bSSatish Balay printf("%d :: in=%d\n",my_id,*in); 1835827bd09bSSatish Balay */ 1836827bd09bSSatish Balay *(buf + *out++) = *(vals + *in++); 1837827bd09bSSatish Balay } 1838827bd09bSSatish Balay 1839827bd09bSSatish Balay grop(buf,work,size,op); 1840827bd09bSSatish Balay 1841827bd09bSSatish Balay in = gs->tree_map_in; 1842827bd09bSSatish Balay out = gs->tree_map_out; 1843827bd09bSSatish Balay 1844827bd09bSSatish Balay while (*in >= 0) 1845827bd09bSSatish Balay {*(vals + *in++) = *(buf + *out++);} 18463fdc5746SBarry Smith PetscFunctionReturn(0); 1847827bd09bSSatish Balay } 1848827bd09bSSatish Balay 1849827bd09bSSatish Balay 1850827bd09bSSatish Balay 1851827bd09bSSatish Balay /****************************************************************************** 1852827bd09bSSatish Balay Function: gather_scatter 1853827bd09bSSatish Balay 1854827bd09bSSatish Balay Input : 1855827bd09bSSatish Balay Output: 1856827bd09bSSatish Balay Return: 1857827bd09bSSatish Balay Description: 1858827bd09bSSatish Balay ******************************************************************************/ 18590924e98cSBarry Smith static PetscErrorCode gs_gop_max( gs_id *gs, PetscScalar *vals) 1860827bd09bSSatish Balay { 18613fdc5746SBarry Smith PetscFunctionBegin; 1862827bd09bSSatish Balay /* local only operations!!! */ 1863827bd09bSSatish Balay if (gs->num_local) 1864827bd09bSSatish Balay {gs_gop_local_max(gs,vals);} 1865827bd09bSSatish Balay 1866827bd09bSSatish Balay /* if intersection tree/pairwise and local isn't empty */ 1867827bd09bSSatish Balay if (gs->num_local_gop) 1868827bd09bSSatish Balay { 1869827bd09bSSatish Balay gs_gop_local_in_max(gs,vals); 1870827bd09bSSatish Balay 1871827bd09bSSatish Balay /* pairwise */ 1872827bd09bSSatish Balay if (gs->num_pairs) 1873827bd09bSSatish Balay {gs_gop_pairwise_max(gs,vals);} 1874827bd09bSSatish Balay 1875827bd09bSSatish Balay /* tree */ 1876827bd09bSSatish Balay else if (gs->max_left_over) 1877827bd09bSSatish Balay {gs_gop_tree_max(gs,vals);} 1878827bd09bSSatish Balay 1879827bd09bSSatish Balay gs_gop_local_out(gs,vals); 1880827bd09bSSatish Balay } 1881827bd09bSSatish Balay /* if intersection tree/pairwise and local is empty */ 1882827bd09bSSatish Balay else 1883827bd09bSSatish Balay { 1884827bd09bSSatish Balay /* pairwise */ 1885827bd09bSSatish Balay if (gs->num_pairs) 1886827bd09bSSatish Balay {gs_gop_pairwise_max(gs,vals);} 1887827bd09bSSatish Balay 1888827bd09bSSatish Balay /* tree */ 1889827bd09bSSatish Balay else if (gs->max_left_over) 1890827bd09bSSatish Balay {gs_gop_tree_max(gs,vals);} 1891827bd09bSSatish Balay } 18923fdc5746SBarry Smith PetscFunctionReturn(0); 1893827bd09bSSatish Balay } 1894827bd09bSSatish Balay 1895827bd09bSSatish Balay 1896827bd09bSSatish Balay 1897827bd09bSSatish Balay /****************************************************************************** 1898827bd09bSSatish Balay Function: gather_scatter 1899827bd09bSSatish Balay 1900827bd09bSSatish Balay Input : 1901827bd09bSSatish Balay Output: 1902827bd09bSSatish Balay Return: 1903827bd09bSSatish Balay Description: 1904827bd09bSSatish Balay ******************************************************************************/ 19050924e98cSBarry Smith static PetscErrorCode gs_gop_local_max( gs_id *gs, PetscScalar *vals) 1906827bd09bSSatish Balay { 1907a501084fSBarry Smith int *num, *map, **reduce; 1908a501084fSBarry Smith PetscScalar tmp; 1909827bd09bSSatish Balay 19103fdc5746SBarry Smith PetscFunctionBegin; 1911827bd09bSSatish Balay num = gs->num_local_reduce; 1912827bd09bSSatish Balay reduce = gs->local_reduce; 1913827bd09bSSatish Balay while ((map = *reduce)) 1914827bd09bSSatish Balay { 1915827bd09bSSatish Balay num ++; 1916827bd09bSSatish Balay tmp = -REAL_MAX; 1917827bd09bSSatish Balay while (*map >= 0) 191839945688SSatish Balay {tmp = PetscMax(tmp,*(vals + *map)); map++;} 1919827bd09bSSatish Balay 1920827bd09bSSatish Balay map = *reduce++; 1921827bd09bSSatish Balay while (*map >= 0) 1922827bd09bSSatish Balay {*(vals + *map++) = tmp;} 1923827bd09bSSatish Balay } 19243fdc5746SBarry Smith PetscFunctionReturn(0); 1925827bd09bSSatish Balay } 1926827bd09bSSatish Balay 1927827bd09bSSatish Balay 1928827bd09bSSatish Balay 1929827bd09bSSatish Balay /****************************************************************************** 1930827bd09bSSatish Balay Function: gather_scatter 1931827bd09bSSatish Balay 1932827bd09bSSatish Balay Input : 1933827bd09bSSatish Balay Output: 1934827bd09bSSatish Balay Return: 1935827bd09bSSatish Balay Description: 1936827bd09bSSatish Balay ******************************************************************************/ 19370924e98cSBarry Smith static PetscErrorCode gs_gop_local_in_max( gs_id *gs, PetscScalar *vals) 1938827bd09bSSatish Balay { 1939a501084fSBarry Smith int *num, *map, **reduce; 1940a501084fSBarry Smith PetscScalar *base; 1941827bd09bSSatish Balay 19423fdc5746SBarry Smith PetscFunctionBegin; 1943827bd09bSSatish Balay num = gs->num_gop_local_reduce; 1944827bd09bSSatish Balay reduce = gs->gop_local_reduce; 1945827bd09bSSatish Balay while ((map = *reduce++)) 1946827bd09bSSatish Balay { 1947827bd09bSSatish Balay num++; 1948827bd09bSSatish Balay base = vals + *map++; 1949827bd09bSSatish Balay while (*map >= 0) 195039945688SSatish Balay {*base = PetscMax(*base,*(vals + *map)); map++;} 1951827bd09bSSatish Balay } 19523fdc5746SBarry Smith PetscFunctionReturn(0); 1953827bd09bSSatish Balay } 1954827bd09bSSatish Balay 1955827bd09bSSatish Balay 1956827bd09bSSatish Balay 1957827bd09bSSatish Balay /****************************************************************************** 1958827bd09bSSatish Balay Function: gather_scatter 1959827bd09bSSatish Balay 1960827bd09bSSatish Balay VERSION 3 :: 1961827bd09bSSatish Balay 1962827bd09bSSatish Balay Input : 1963827bd09bSSatish Balay Output: 1964827bd09bSSatish Balay Return: 1965827bd09bSSatish Balay Description: 1966827bd09bSSatish Balay ******************************************************************************/ 19670924e98cSBarry Smith static PetscErrorCode gs_gop_pairwise_max( gs_id *gs, PetscScalar *in_vals) 1968827bd09bSSatish Balay { 1969a501084fSBarry Smith PetscScalar *dptr1, *dptr2, *dptr3, *in1, *in2; 1970a501084fSBarry Smith int *iptr, *msg_list, *msg_size, **msg_nodes; 1971a501084fSBarry Smith int *pw, *list, *size, **nodes; 1972827bd09bSSatish Balay MPI_Request *msg_ids_in, *msg_ids_out, *ids_in, *ids_out; 1973827bd09bSSatish Balay MPI_Status status; 19743fdc5746SBarry Smith PetscErrorCode ierr; 1975827bd09bSSatish Balay 19763fdc5746SBarry Smith PetscFunctionBegin; 1977a501084fSBarry Smith /* strip and load s */ 1978827bd09bSSatish Balay msg_list =list = gs->pair_list; 1979827bd09bSSatish Balay msg_size =size = gs->msg_sizes; 1980827bd09bSSatish Balay msg_nodes=nodes = gs->node_list; 1981827bd09bSSatish Balay iptr=pw = gs->pw_elm_list; 1982827bd09bSSatish Balay dptr1=dptr3 = gs->pw_vals; 1983827bd09bSSatish Balay msg_ids_in = ids_in = gs->msg_ids_in; 1984827bd09bSSatish Balay msg_ids_out = ids_out = gs->msg_ids_out; 1985827bd09bSSatish Balay dptr2 = gs->out; 1986827bd09bSSatish Balay in1=in2 = gs->in; 1987827bd09bSSatish Balay 1988827bd09bSSatish Balay /* post the receives */ 1989827bd09bSSatish Balay /* msg_nodes=nodes; */ 1990827bd09bSSatish Balay do 1991827bd09bSSatish Balay { 1992827bd09bSSatish Balay /* Should MPI_ANY_SOURCE be replaced by *list ? In that case do the 1993827bd09bSSatish Balay second one *list and do list++ afterwards */ 19943fdc5746SBarry Smith ierr = MPI_Irecv(in1, *size, MPIU_SCALAR, MPI_ANY_SOURCE, MSGTAG1 + *list++, gs->gs_comm, msg_ids_in++);CHKERRQ(ierr); 1995827bd09bSSatish Balay in1 += *size++; 1996827bd09bSSatish Balay } 1997827bd09bSSatish Balay while (*++msg_nodes); 1998827bd09bSSatish Balay msg_nodes=nodes; 1999827bd09bSSatish Balay 2000827bd09bSSatish Balay /* load gs values into in out gs buffers */ 2001827bd09bSSatish Balay while (*iptr >= 0) 2002827bd09bSSatish Balay {*dptr3++ = *(in_vals + *iptr++);} 2003827bd09bSSatish Balay 2004827bd09bSSatish Balay /* load out buffers and post the sends */ 2005827bd09bSSatish Balay while ((iptr = *msg_nodes++)) 2006827bd09bSSatish Balay { 2007827bd09bSSatish Balay dptr3 = dptr2; 2008827bd09bSSatish Balay while (*iptr >= 0) 2009827bd09bSSatish Balay {*dptr2++ = *(dptr1 + *iptr++);} 2010827bd09bSSatish Balay /* CHECK PERSISTENT COMMS MODE FOR ALL THIS STUFF */ 2011827bd09bSSatish Balay /* is msg_ids_out++ correct? */ 20123fdc5746SBarry Smith ierr = MPI_Isend(dptr3, *msg_size++, MPIU_SCALAR, *msg_list++, MSGTAG1+my_id, gs->gs_comm, msg_ids_out++);CHKERRQ(ierr); 2013827bd09bSSatish Balay } 2014827bd09bSSatish Balay 2015827bd09bSSatish Balay if (gs->max_left_over) 2016827bd09bSSatish Balay {gs_gop_tree_max(gs,in_vals);} 2017827bd09bSSatish Balay 2018827bd09bSSatish Balay /* process the received data */ 2019827bd09bSSatish Balay msg_nodes=nodes; 2020827bd09bSSatish Balay while ((iptr = *nodes++)) 2021827bd09bSSatish Balay { 2022827bd09bSSatish Balay /* Should I check the return value of MPI_Wait() or status? */ 2023827bd09bSSatish Balay /* Can this loop be replaced by a call to MPI_Waitall()? */ 20243fdc5746SBarry Smith ierr = MPI_Wait(ids_in++, &status);CHKERRQ(ierr); 2025827bd09bSSatish Balay while (*iptr >= 0) 202639945688SSatish Balay {*(dptr1 + *iptr) = PetscMax(*(dptr1 + *iptr),*in2); iptr++; in2++;} 2027827bd09bSSatish Balay } 2028827bd09bSSatish Balay 2029827bd09bSSatish Balay /* replace vals */ 2030827bd09bSSatish Balay while (*pw >= 0) 2031827bd09bSSatish Balay {*(in_vals + *pw++) = *dptr1++;} 2032827bd09bSSatish Balay 2033827bd09bSSatish Balay /* clear isend message handles */ 2034827bd09bSSatish Balay /* This changed for clarity though it could be the same */ 2035827bd09bSSatish Balay while (*msg_nodes++) 2036827bd09bSSatish Balay /* Should I check the return value of MPI_Wait() or status? */ 2037827bd09bSSatish Balay /* Can this loop be replaced by a call to MPI_Waitall()? */ 20383fdc5746SBarry Smith {ierr = MPI_Wait(ids_out++, &status);CHKERRQ(ierr);} 20393fdc5746SBarry Smith PetscFunctionReturn(0); 2040827bd09bSSatish Balay } 2041827bd09bSSatish Balay 2042827bd09bSSatish Balay 2043827bd09bSSatish Balay 2044827bd09bSSatish Balay /****************************************************************************** 2045827bd09bSSatish Balay Function: gather_scatter 2046827bd09bSSatish Balay 2047827bd09bSSatish Balay Input : 2048827bd09bSSatish Balay Output: 2049827bd09bSSatish Balay Return: 2050827bd09bSSatish Balay Description: 2051827bd09bSSatish Balay ******************************************************************************/ 20520924e98cSBarry Smith static PetscErrorCode gs_gop_tree_max(gs_id *gs, PetscScalar *vals) 2053827bd09bSSatish Balay { 2054827bd09bSSatish Balay int size; 2055827bd09bSSatish Balay int *in, *out; 2056a501084fSBarry Smith PetscScalar *buf, *work; 20573fdc5746SBarry Smith PetscErrorCode ierr; 2058827bd09bSSatish Balay 20593fdc5746SBarry Smith PetscFunctionBegin; 2060827bd09bSSatish Balay in = gs->tree_map_in; 2061827bd09bSSatish Balay out = gs->tree_map_out; 2062827bd09bSSatish Balay buf = gs->tree_buf; 2063827bd09bSSatish Balay work = gs->tree_work; 2064827bd09bSSatish Balay size = gs->tree_nel; 2065827bd09bSSatish Balay 2066827bd09bSSatish Balay rvec_set(buf,-REAL_MAX,size); 2067827bd09bSSatish Balay 2068827bd09bSSatish Balay while (*in >= 0) 2069827bd09bSSatish Balay {*(buf + *out++) = *(vals + *in++);} 2070827bd09bSSatish Balay 2071827bd09bSSatish Balay in = gs->tree_map_in; 2072827bd09bSSatish Balay out = gs->tree_map_out; 20733fdc5746SBarry Smith ierr = MPI_Allreduce(buf,work,size,MPIU_SCALAR,MPI_MAX,gs->gs_comm);CHKERRQ(ierr); 2074827bd09bSSatish Balay while (*in >= 0) 2075827bd09bSSatish Balay {*(vals + *in++) = *(work + *out++);} 20763fdc5746SBarry Smith PetscFunctionReturn(0); 2077827bd09bSSatish Balay } 2078827bd09bSSatish Balay 2079827bd09bSSatish Balay 2080827bd09bSSatish Balay 2081827bd09bSSatish Balay /****************************************************************************** 2082827bd09bSSatish Balay Function: gather_scatter 2083827bd09bSSatish Balay 2084827bd09bSSatish Balay Input : 2085827bd09bSSatish Balay Output: 2086827bd09bSSatish Balay Return: 2087827bd09bSSatish Balay Description: 2088827bd09bSSatish Balay ******************************************************************************/ 20890924e98cSBarry Smith static PetscErrorCode gs_gop_min_abs( gs_id *gs, PetscScalar *vals) 2090827bd09bSSatish Balay { 20913fdc5746SBarry Smith PetscFunctionBegin; 2092827bd09bSSatish Balay /* local only operations!!! */ 2093827bd09bSSatish Balay if (gs->num_local) 2094827bd09bSSatish Balay {gs_gop_local_min_abs(gs,vals);} 2095827bd09bSSatish Balay 2096827bd09bSSatish Balay /* if intersection tree/pairwise and local isn't empty */ 2097827bd09bSSatish Balay if (gs->num_local_gop) 2098827bd09bSSatish Balay { 2099827bd09bSSatish Balay gs_gop_local_in_min_abs(gs,vals); 2100827bd09bSSatish Balay 2101827bd09bSSatish Balay /* pairwise */ 2102827bd09bSSatish Balay if (gs->num_pairs) 2103827bd09bSSatish Balay {gs_gop_pairwise_min_abs(gs,vals);} 2104827bd09bSSatish Balay 2105827bd09bSSatish Balay /* tree */ 2106827bd09bSSatish Balay else if (gs->max_left_over) 2107827bd09bSSatish Balay {gs_gop_tree_min_abs(gs,vals);} 2108827bd09bSSatish Balay 2109827bd09bSSatish Balay gs_gop_local_out(gs,vals); 2110827bd09bSSatish Balay } 2111827bd09bSSatish Balay /* if intersection tree/pairwise and local is empty */ 2112827bd09bSSatish Balay else 2113827bd09bSSatish Balay { 2114827bd09bSSatish Balay /* pairwise */ 2115827bd09bSSatish Balay if (gs->num_pairs) 2116827bd09bSSatish Balay {gs_gop_pairwise_min_abs(gs,vals);} 2117827bd09bSSatish Balay 2118827bd09bSSatish Balay /* tree */ 2119827bd09bSSatish Balay else if (gs->max_left_over) 2120827bd09bSSatish Balay {gs_gop_tree_min_abs(gs,vals);} 2121827bd09bSSatish Balay } 21223fdc5746SBarry Smith PetscFunctionReturn(0); 2123827bd09bSSatish Balay } 2124827bd09bSSatish Balay 2125827bd09bSSatish Balay 2126827bd09bSSatish Balay 2127827bd09bSSatish Balay /****************************************************************************** 2128827bd09bSSatish Balay Function: gather_scatter 2129827bd09bSSatish Balay 2130827bd09bSSatish Balay Input : 2131827bd09bSSatish Balay Output: 2132827bd09bSSatish Balay Return: 2133827bd09bSSatish Balay Description: 2134827bd09bSSatish Balay ******************************************************************************/ 21350924e98cSBarry Smith static PetscErrorCode gs_gop_local_min_abs( gs_id *gs, PetscScalar *vals) 2136827bd09bSSatish Balay { 2137a501084fSBarry Smith int *num, *map, **reduce; 2138a501084fSBarry Smith PetscScalar tmp; 2139827bd09bSSatish Balay 21403fdc5746SBarry Smith PetscFunctionBegin; 2141827bd09bSSatish Balay num = gs->num_local_reduce; 2142827bd09bSSatish Balay reduce = gs->local_reduce; 2143827bd09bSSatish Balay while ((map = *reduce)) 2144827bd09bSSatish Balay { 2145827bd09bSSatish Balay num ++; 2146827bd09bSSatish Balay tmp = REAL_MAX; 2147827bd09bSSatish Balay while (*map >= 0) 2148827bd09bSSatish Balay {tmp = MIN_FABS(tmp,*(vals + *map)); map++;} 2149827bd09bSSatish Balay 2150827bd09bSSatish Balay map = *reduce++; 2151827bd09bSSatish Balay while (*map >= 0) 2152827bd09bSSatish Balay {*(vals + *map++) = tmp;} 2153827bd09bSSatish Balay } 21543fdc5746SBarry Smith PetscFunctionReturn(0); 2155827bd09bSSatish Balay } 2156827bd09bSSatish Balay 2157827bd09bSSatish Balay 2158827bd09bSSatish Balay 2159827bd09bSSatish Balay /****************************************************************************** 2160827bd09bSSatish Balay Function: gather_scatter 2161827bd09bSSatish Balay 2162827bd09bSSatish Balay Input : 2163827bd09bSSatish Balay Output: 2164827bd09bSSatish Balay Return: 2165827bd09bSSatish Balay Description: 2166827bd09bSSatish Balay ******************************************************************************/ 21670924e98cSBarry Smith static PetscErrorCode gs_gop_local_in_min_abs( gs_id *gs, PetscScalar *vals) 2168827bd09bSSatish Balay { 2169a501084fSBarry Smith int *num, *map, **reduce; 2170a501084fSBarry Smith PetscScalar *base; 2171827bd09bSSatish Balay 21723fdc5746SBarry Smith PetscFunctionBegin; 2173827bd09bSSatish Balay num = gs->num_gop_local_reduce; 2174827bd09bSSatish Balay reduce = gs->gop_local_reduce; 2175827bd09bSSatish Balay while ((map = *reduce++)) 2176827bd09bSSatish Balay { 2177827bd09bSSatish Balay num++; 2178827bd09bSSatish Balay base = vals + *map++; 2179827bd09bSSatish Balay while (*map >= 0) 2180827bd09bSSatish Balay {*base = MIN_FABS(*base,*(vals + *map)); map++;} 2181827bd09bSSatish Balay } 21823fdc5746SBarry Smith PetscFunctionReturn(0); 2183827bd09bSSatish Balay } 2184827bd09bSSatish Balay 2185827bd09bSSatish Balay 2186827bd09bSSatish Balay 2187827bd09bSSatish Balay /****************************************************************************** 2188827bd09bSSatish Balay Function: gather_scatter 2189827bd09bSSatish Balay 2190827bd09bSSatish Balay VERSION 3 :: 2191827bd09bSSatish Balay 2192827bd09bSSatish Balay Input : 2193827bd09bSSatish Balay Output: 2194827bd09bSSatish Balay Return: 2195827bd09bSSatish Balay Description: 2196827bd09bSSatish Balay ******************************************************************************/ 21970924e98cSBarry Smith static PetscErrorCode gs_gop_pairwise_min_abs( gs_id *gs, PetscScalar *in_vals) 2198827bd09bSSatish Balay { 2199a501084fSBarry Smith PetscScalar *dptr1, *dptr2, *dptr3, *in1, *in2; 2200a501084fSBarry Smith int *iptr, *msg_list, *msg_size, **msg_nodes; 2201a501084fSBarry Smith int *pw, *list, *size, **nodes; 2202827bd09bSSatish Balay MPI_Request *msg_ids_in, *msg_ids_out, *ids_in, *ids_out; 2203827bd09bSSatish Balay MPI_Status status; 22043fdc5746SBarry Smith PetscErrorCode ierr; 2205827bd09bSSatish Balay 22063fdc5746SBarry Smith PetscFunctionBegin; 2207a501084fSBarry Smith /* strip and load s */ 2208827bd09bSSatish Balay msg_list =list = gs->pair_list; 2209827bd09bSSatish Balay msg_size =size = gs->msg_sizes; 2210827bd09bSSatish Balay msg_nodes=nodes = gs->node_list; 2211827bd09bSSatish Balay iptr=pw = gs->pw_elm_list; 2212827bd09bSSatish Balay dptr1=dptr3 = gs->pw_vals; 2213827bd09bSSatish Balay msg_ids_in = ids_in = gs->msg_ids_in; 2214827bd09bSSatish Balay msg_ids_out = ids_out = gs->msg_ids_out; 2215827bd09bSSatish Balay dptr2 = gs->out; 2216827bd09bSSatish Balay in1=in2 = gs->in; 2217827bd09bSSatish Balay 2218827bd09bSSatish Balay /* post the receives */ 2219827bd09bSSatish Balay /* msg_nodes=nodes; */ 2220827bd09bSSatish Balay do 2221827bd09bSSatish Balay { 2222827bd09bSSatish Balay /* Should MPI_ANY_SOURCE be replaced by *list ? In that case do the 2223827bd09bSSatish Balay second one *list and do list++ afterwards */ 22243fdc5746SBarry Smith ierr = MPI_Irecv(in1, *size, MPIU_SCALAR, MPI_ANY_SOURCE, MSGTAG1 + *list++, gs->gs_comm, msg_ids_in++);CHKERRQ(ierr); 2225827bd09bSSatish Balay in1 += *size++; 2226827bd09bSSatish Balay } 2227827bd09bSSatish Balay while (*++msg_nodes); 2228827bd09bSSatish Balay msg_nodes=nodes; 2229827bd09bSSatish Balay 2230827bd09bSSatish Balay /* load gs values into in out gs buffers */ 2231827bd09bSSatish Balay while (*iptr >= 0) 2232827bd09bSSatish Balay {*dptr3++ = *(in_vals + *iptr++);} 2233827bd09bSSatish Balay 2234827bd09bSSatish Balay /* load out buffers and post the sends */ 2235827bd09bSSatish Balay while ((iptr = *msg_nodes++)) 2236827bd09bSSatish Balay { 2237827bd09bSSatish Balay dptr3 = dptr2; 2238827bd09bSSatish Balay while (*iptr >= 0) 2239827bd09bSSatish Balay {*dptr2++ = *(dptr1 + *iptr++);} 2240827bd09bSSatish Balay /* CHECK PERSISTENT COMMS MODE FOR ALL THIS STUFF */ 2241827bd09bSSatish Balay /* is msg_ids_out++ correct? */ 22423fdc5746SBarry Smith ierr = MPI_Isend(dptr3, *msg_size++, MPIU_SCALAR, *msg_list++, MSGTAG1+my_id, gs->gs_comm, msg_ids_out++);CHKERRQ(ierr); 2243827bd09bSSatish Balay } 2244827bd09bSSatish Balay 2245827bd09bSSatish Balay if (gs->max_left_over) 2246827bd09bSSatish Balay {gs_gop_tree_min_abs(gs,in_vals);} 2247827bd09bSSatish Balay 2248827bd09bSSatish Balay /* process the received data */ 2249827bd09bSSatish Balay msg_nodes=nodes; 2250827bd09bSSatish Balay while ((iptr = *nodes++)) 2251827bd09bSSatish Balay { 2252827bd09bSSatish Balay /* Should I check the return value of MPI_Wait() or status? */ 2253827bd09bSSatish Balay /* Can this loop be replaced by a call to MPI_Waitall()? */ 22543fdc5746SBarry Smith ierr = MPI_Wait(ids_in++, &status);CHKERRQ(ierr); 2255827bd09bSSatish Balay while (*iptr >= 0) 2256827bd09bSSatish Balay {*(dptr1 + *iptr) = MIN_FABS(*(dptr1 + *iptr),*in2); iptr++; in2++;} 2257827bd09bSSatish Balay } 2258827bd09bSSatish Balay 2259827bd09bSSatish Balay /* replace vals */ 2260827bd09bSSatish Balay while (*pw >= 0) 2261827bd09bSSatish Balay {*(in_vals + *pw++) = *dptr1++;} 2262827bd09bSSatish Balay 2263827bd09bSSatish Balay /* clear isend message handles */ 2264827bd09bSSatish Balay /* This changed for clarity though it could be the same */ 2265827bd09bSSatish Balay while (*msg_nodes++) 2266827bd09bSSatish Balay /* Should I check the return value of MPI_Wait() or status? */ 2267827bd09bSSatish Balay /* Can this loop be replaced by a call to MPI_Waitall()? */ 22683fdc5746SBarry Smith {ierr = MPI_Wait(ids_out++, &status);CHKERRQ(ierr);} 22693fdc5746SBarry Smith PetscFunctionReturn(0); 2270827bd09bSSatish Balay } 2271827bd09bSSatish Balay 2272827bd09bSSatish Balay 2273827bd09bSSatish Balay 2274827bd09bSSatish Balay /****************************************************************************** 2275827bd09bSSatish Balay Function: gather_scatter 2276827bd09bSSatish Balay 2277827bd09bSSatish Balay Input : 2278827bd09bSSatish Balay Output: 2279827bd09bSSatish Balay Return: 2280827bd09bSSatish Balay Description: 2281827bd09bSSatish Balay ******************************************************************************/ 22820924e98cSBarry Smith static PetscErrorCode gs_gop_tree_min_abs(gs_id *gs, PetscScalar *vals) 2283827bd09bSSatish Balay { 2284827bd09bSSatish Balay int size; 2285827bd09bSSatish Balay int *in, *out; 2286a501084fSBarry Smith PetscScalar *buf, *work; 2287827bd09bSSatish Balay int op[] = {GL_MIN_ABS,0}; 2288827bd09bSSatish Balay 22893fdc5746SBarry Smith PetscFunctionBegin; 2290827bd09bSSatish Balay in = gs->tree_map_in; 2291827bd09bSSatish Balay out = gs->tree_map_out; 2292827bd09bSSatish Balay buf = gs->tree_buf; 2293827bd09bSSatish Balay work = gs->tree_work; 2294827bd09bSSatish Balay size = gs->tree_nel; 2295827bd09bSSatish Balay 2296827bd09bSSatish Balay rvec_set(buf,REAL_MAX,size); 2297827bd09bSSatish Balay 2298827bd09bSSatish Balay while (*in >= 0) 2299827bd09bSSatish Balay {*(buf + *out++) = *(vals + *in++);} 2300827bd09bSSatish Balay 2301827bd09bSSatish Balay in = gs->tree_map_in; 2302827bd09bSSatish Balay out = gs->tree_map_out; 2303827bd09bSSatish Balay grop(buf,work,size,op); 2304827bd09bSSatish Balay while (*in >= 0) 2305827bd09bSSatish Balay {*(vals + *in++) = *(buf + *out++);} 23063fdc5746SBarry Smith PetscFunctionReturn(0); 2307827bd09bSSatish Balay } 2308827bd09bSSatish Balay 2309827bd09bSSatish Balay 2310827bd09bSSatish Balay 2311827bd09bSSatish Balay /****************************************************************************** 2312827bd09bSSatish Balay Function: gather_scatter 2313827bd09bSSatish Balay 2314827bd09bSSatish Balay Input : 2315827bd09bSSatish Balay Output: 2316827bd09bSSatish Balay Return: 2317827bd09bSSatish Balay Description: 2318827bd09bSSatish Balay ******************************************************************************/ 23190924e98cSBarry Smith static PetscErrorCode gs_gop_min( gs_id *gs, PetscScalar *vals) 2320827bd09bSSatish Balay { 23213fdc5746SBarry Smith PetscFunctionBegin; 2322827bd09bSSatish Balay /* local only operations!!! */ 2323827bd09bSSatish Balay if (gs->num_local) 2324827bd09bSSatish Balay {gs_gop_local_min(gs,vals);} 2325827bd09bSSatish Balay 2326827bd09bSSatish Balay /* if intersection tree/pairwise and local isn't empty */ 2327827bd09bSSatish Balay if (gs->num_local_gop) 2328827bd09bSSatish Balay { 2329827bd09bSSatish Balay gs_gop_local_in_min(gs,vals); 2330827bd09bSSatish Balay 2331827bd09bSSatish Balay /* pairwise */ 2332827bd09bSSatish Balay if (gs->num_pairs) 2333827bd09bSSatish Balay {gs_gop_pairwise_min(gs,vals);} 2334827bd09bSSatish Balay 2335827bd09bSSatish Balay /* tree */ 2336827bd09bSSatish Balay else if (gs->max_left_over) 2337827bd09bSSatish Balay {gs_gop_tree_min(gs,vals);} 2338827bd09bSSatish Balay 2339827bd09bSSatish Balay gs_gop_local_out(gs,vals); 2340827bd09bSSatish Balay } 2341827bd09bSSatish Balay /* if intersection tree/pairwise and local is empty */ 2342827bd09bSSatish Balay else 2343827bd09bSSatish Balay { 2344827bd09bSSatish Balay /* pairwise */ 2345827bd09bSSatish Balay if (gs->num_pairs) 2346827bd09bSSatish Balay {gs_gop_pairwise_min(gs,vals);} 2347827bd09bSSatish Balay 2348827bd09bSSatish Balay /* tree */ 2349827bd09bSSatish Balay else if (gs->max_left_over) 2350827bd09bSSatish Balay {gs_gop_tree_min(gs,vals);} 2351827bd09bSSatish Balay } 23523fdc5746SBarry Smith PetscFunctionReturn(0); 2353827bd09bSSatish Balay } 2354827bd09bSSatish Balay 2355827bd09bSSatish Balay 2356827bd09bSSatish Balay 2357827bd09bSSatish Balay /****************************************************************************** 2358827bd09bSSatish Balay Function: gather_scatter 2359827bd09bSSatish Balay 2360827bd09bSSatish Balay Input : 2361827bd09bSSatish Balay Output: 2362827bd09bSSatish Balay Return: 2363827bd09bSSatish Balay Description: 2364827bd09bSSatish Balay ******************************************************************************/ 23650924e98cSBarry Smith static PetscErrorCode gs_gop_local_min( gs_id *gs, PetscScalar *vals) 2366827bd09bSSatish Balay { 2367a501084fSBarry Smith int *num, *map, **reduce; 2368a501084fSBarry Smith PetscScalar tmp; 23693fdc5746SBarry Smith PetscFunctionBegin; 2370827bd09bSSatish Balay num = gs->num_local_reduce; 2371827bd09bSSatish Balay reduce = gs->local_reduce; 2372827bd09bSSatish Balay while ((map = *reduce)) 2373827bd09bSSatish Balay { 2374827bd09bSSatish Balay num ++; 2375827bd09bSSatish Balay tmp = REAL_MAX; 2376827bd09bSSatish Balay while (*map >= 0) 237739945688SSatish Balay {tmp = PetscMin(tmp,*(vals + *map)); map++;} 2378827bd09bSSatish Balay 2379827bd09bSSatish Balay map = *reduce++; 2380827bd09bSSatish Balay while (*map >= 0) 2381827bd09bSSatish Balay {*(vals + *map++) = tmp;} 2382827bd09bSSatish Balay } 23833fdc5746SBarry Smith PetscFunctionReturn(0); 2384827bd09bSSatish Balay } 2385827bd09bSSatish Balay 2386827bd09bSSatish Balay 2387827bd09bSSatish Balay 2388827bd09bSSatish Balay /****************************************************************************** 2389827bd09bSSatish Balay Function: gather_scatter 2390827bd09bSSatish Balay 2391827bd09bSSatish Balay Input : 2392827bd09bSSatish Balay Output: 2393827bd09bSSatish Balay Return: 2394827bd09bSSatish Balay Description: 2395827bd09bSSatish Balay ******************************************************************************/ 23960924e98cSBarry Smith static PetscErrorCode gs_gop_local_in_min( gs_id *gs, PetscScalar *vals) 2397827bd09bSSatish Balay { 2398a501084fSBarry Smith int *num, *map, **reduce; 2399a501084fSBarry Smith PetscScalar *base; 2400827bd09bSSatish Balay 24013fdc5746SBarry Smith PetscFunctionBegin; 2402827bd09bSSatish Balay num = gs->num_gop_local_reduce; 2403827bd09bSSatish Balay reduce = gs->gop_local_reduce; 2404827bd09bSSatish Balay while ((map = *reduce++)) 2405827bd09bSSatish Balay { 2406827bd09bSSatish Balay num++; 2407827bd09bSSatish Balay base = vals + *map++; 2408827bd09bSSatish Balay while (*map >= 0) 240939945688SSatish Balay {*base = PetscMin(*base,*(vals + *map)); map++;} 2410827bd09bSSatish Balay } 24113fdc5746SBarry Smith PetscFunctionReturn(0); 2412827bd09bSSatish Balay } 2413827bd09bSSatish Balay 2414827bd09bSSatish Balay 2415827bd09bSSatish Balay 2416827bd09bSSatish Balay /****************************************************************************** 2417827bd09bSSatish Balay Function: gather_scatter 2418827bd09bSSatish Balay 2419827bd09bSSatish Balay VERSION 3 :: 2420827bd09bSSatish Balay 2421827bd09bSSatish Balay Input : 2422827bd09bSSatish Balay Output: 2423827bd09bSSatish Balay Return: 2424827bd09bSSatish Balay Description: 2425827bd09bSSatish Balay ******************************************************************************/ 24260924e98cSBarry Smith static PetscErrorCode gs_gop_pairwise_min( gs_id *gs, PetscScalar *in_vals) 2427827bd09bSSatish Balay { 2428a501084fSBarry Smith PetscScalar *dptr1, *dptr2, *dptr3, *in1, *in2; 2429a501084fSBarry Smith int *iptr, *msg_list, *msg_size, **msg_nodes; 2430a501084fSBarry Smith int *pw, *list, *size, **nodes; 2431827bd09bSSatish Balay MPI_Request *msg_ids_in, *msg_ids_out, *ids_in, *ids_out; 2432827bd09bSSatish Balay MPI_Status status; 24333fdc5746SBarry Smith PetscErrorCode ierr; 2434827bd09bSSatish Balay 24353fdc5746SBarry Smith PetscFunctionBegin; 2436a501084fSBarry Smith /* strip and load s */ 2437827bd09bSSatish Balay msg_list =list = gs->pair_list; 2438827bd09bSSatish Balay msg_size =size = gs->msg_sizes; 2439827bd09bSSatish Balay msg_nodes=nodes = gs->node_list; 2440827bd09bSSatish Balay iptr=pw = gs->pw_elm_list; 2441827bd09bSSatish Balay dptr1=dptr3 = gs->pw_vals; 2442827bd09bSSatish Balay msg_ids_in = ids_in = gs->msg_ids_in; 2443827bd09bSSatish Balay msg_ids_out = ids_out = gs->msg_ids_out; 2444827bd09bSSatish Balay dptr2 = gs->out; 2445827bd09bSSatish Balay in1=in2 = gs->in; 2446827bd09bSSatish Balay 2447827bd09bSSatish Balay /* post the receives */ 2448827bd09bSSatish Balay /* msg_nodes=nodes; */ 2449827bd09bSSatish Balay do 2450827bd09bSSatish Balay { 2451827bd09bSSatish Balay /* Should MPI_ANY_SOURCE be replaced by *list ? In that case do the 2452827bd09bSSatish Balay second one *list and do list++ afterwards */ 24533fdc5746SBarry Smith ierr = MPI_Irecv(in1, *size, MPIU_SCALAR, MPI_ANY_SOURCE, MSGTAG1 + *list++, gs->gs_comm, msg_ids_in++);CHKERRQ(ierr); 2454827bd09bSSatish Balay in1 += *size++; 2455827bd09bSSatish Balay } 2456827bd09bSSatish Balay while (*++msg_nodes); 2457827bd09bSSatish Balay msg_nodes=nodes; 2458827bd09bSSatish Balay 2459827bd09bSSatish Balay /* load gs values into in out gs buffers */ 2460827bd09bSSatish Balay while (*iptr >= 0) 2461827bd09bSSatish Balay {*dptr3++ = *(in_vals + *iptr++);} 2462827bd09bSSatish Balay 2463827bd09bSSatish Balay /* load out buffers and post the sends */ 2464827bd09bSSatish Balay while ((iptr = *msg_nodes++)) 2465827bd09bSSatish Balay { 2466827bd09bSSatish Balay dptr3 = dptr2; 2467827bd09bSSatish Balay while (*iptr >= 0) 2468827bd09bSSatish Balay {*dptr2++ = *(dptr1 + *iptr++);} 2469827bd09bSSatish Balay /* CHECK PERSISTENT COMMS MODE FOR ALL THIS STUFF */ 2470827bd09bSSatish Balay /* is msg_ids_out++ correct? */ 24713fdc5746SBarry Smith ierr = MPI_Isend(dptr3, *msg_size++, MPIU_SCALAR, *msg_list++, MSGTAG1+my_id, gs->gs_comm, msg_ids_out++);CHKERRQ(ierr); 2472827bd09bSSatish Balay } 2473827bd09bSSatish Balay 2474827bd09bSSatish Balay /* process the received data */ 2475827bd09bSSatish Balay if (gs->max_left_over) 2476827bd09bSSatish Balay {gs_gop_tree_min(gs,in_vals);} 2477827bd09bSSatish Balay 2478827bd09bSSatish Balay msg_nodes=nodes; 2479827bd09bSSatish Balay while ((iptr = *nodes++)) 2480827bd09bSSatish Balay { 2481827bd09bSSatish Balay /* Should I check the return value of MPI_Wait() or status? */ 2482827bd09bSSatish Balay /* Can this loop be replaced by a call to MPI_Waitall()? */ 24833fdc5746SBarry Smith ierr = MPI_Wait(ids_in++, &status);CHKERRQ(ierr); 2484827bd09bSSatish Balay while (*iptr >= 0) 248539945688SSatish Balay {*(dptr1 + *iptr) = PetscMin(*(dptr1 + *iptr),*in2); iptr++; in2++;} 2486827bd09bSSatish Balay } 2487827bd09bSSatish Balay 2488827bd09bSSatish Balay /* replace vals */ 2489827bd09bSSatish Balay while (*pw >= 0) 2490827bd09bSSatish Balay {*(in_vals + *pw++) = *dptr1++;} 2491827bd09bSSatish Balay 2492827bd09bSSatish Balay /* clear isend message handles */ 2493827bd09bSSatish Balay /* This changed for clarity though it could be the same */ 2494827bd09bSSatish Balay while (*msg_nodes++) 2495827bd09bSSatish Balay /* Should I check the return value of MPI_Wait() or status? */ 2496827bd09bSSatish Balay /* Can this loop be replaced by a call to MPI_Waitall()? */ 24973fdc5746SBarry Smith {ierr = MPI_Wait(ids_out++, &status);CHKERRQ(ierr);} 24983fdc5746SBarry Smith PetscFunctionReturn(0); 2499827bd09bSSatish Balay } 2500827bd09bSSatish Balay 2501827bd09bSSatish Balay 2502827bd09bSSatish Balay 2503827bd09bSSatish Balay /****************************************************************************** 2504827bd09bSSatish Balay Function: gather_scatter 2505827bd09bSSatish Balay 2506827bd09bSSatish Balay Input : 2507827bd09bSSatish Balay Output: 2508827bd09bSSatish Balay Return: 2509827bd09bSSatish Balay Description: 2510827bd09bSSatish Balay ******************************************************************************/ 25110924e98cSBarry Smith static PetscErrorCode gs_gop_tree_min(gs_id *gs, PetscScalar *vals) 2512827bd09bSSatish Balay { 2513827bd09bSSatish Balay int size; 2514827bd09bSSatish Balay int *in, *out; 2515a501084fSBarry Smith PetscScalar *buf, *work; 25163fdc5746SBarry Smith PetscErrorCode ierr; 2517827bd09bSSatish Balay 25183fdc5746SBarry Smith PetscFunctionBegin; 2519827bd09bSSatish Balay in = gs->tree_map_in; 2520827bd09bSSatish Balay out = gs->tree_map_out; 2521827bd09bSSatish Balay buf = gs->tree_buf; 2522827bd09bSSatish Balay work = gs->tree_work; 2523827bd09bSSatish Balay size = gs->tree_nel; 2524827bd09bSSatish Balay 2525827bd09bSSatish Balay rvec_set(buf,REAL_MAX,size); 2526827bd09bSSatish Balay 2527827bd09bSSatish Balay while (*in >= 0) 2528827bd09bSSatish Balay {*(buf + *out++) = *(vals + *in++);} 2529827bd09bSSatish Balay 2530827bd09bSSatish Balay in = gs->tree_map_in; 2531827bd09bSSatish Balay out = gs->tree_map_out; 25323fdc5746SBarry Smith ierr = MPI_Allreduce(buf,work,size,MPIU_SCALAR,MPI_MIN,gs->gs_comm);CHKERRQ(ierr); 2533827bd09bSSatish Balay while (*in >= 0) 2534827bd09bSSatish Balay {*(vals + *in++) = *(work + *out++);} 25353fdc5746SBarry Smith PetscFunctionReturn(0); 2536827bd09bSSatish Balay } 2537827bd09bSSatish Balay 2538827bd09bSSatish Balay 2539827bd09bSSatish Balay 2540827bd09bSSatish Balay /****************************************************************************** 2541827bd09bSSatish Balay Function: gather_scatter 2542827bd09bSSatish Balay 2543827bd09bSSatish Balay Input : 2544827bd09bSSatish Balay Output: 2545827bd09bSSatish Balay Return: 2546827bd09bSSatish Balay Description: 2547827bd09bSSatish Balay ******************************************************************************/ 25480924e98cSBarry Smith static PetscErrorCode gs_gop_times( gs_id *gs, PetscScalar *vals) 2549827bd09bSSatish Balay { 25503fdc5746SBarry Smith PetscFunctionBegin; 2551827bd09bSSatish Balay /* local only operations!!! */ 2552827bd09bSSatish Balay if (gs->num_local) 2553827bd09bSSatish Balay {gs_gop_local_times(gs,vals);} 2554827bd09bSSatish Balay 2555827bd09bSSatish Balay /* if intersection tree/pairwise and local isn't empty */ 2556827bd09bSSatish Balay if (gs->num_local_gop) 2557827bd09bSSatish Balay { 2558827bd09bSSatish Balay gs_gop_local_in_times(gs,vals); 2559827bd09bSSatish Balay 2560827bd09bSSatish Balay /* pairwise */ 2561827bd09bSSatish Balay if (gs->num_pairs) 2562827bd09bSSatish Balay {gs_gop_pairwise_times(gs,vals);} 2563827bd09bSSatish Balay 2564827bd09bSSatish Balay /* tree */ 2565827bd09bSSatish Balay else if (gs->max_left_over) 2566827bd09bSSatish Balay {gs_gop_tree_times(gs,vals);} 2567827bd09bSSatish Balay 2568827bd09bSSatish Balay gs_gop_local_out(gs,vals); 2569827bd09bSSatish Balay } 2570827bd09bSSatish Balay /* if intersection tree/pairwise and local is empty */ 2571827bd09bSSatish Balay else 2572827bd09bSSatish Balay { 2573827bd09bSSatish Balay /* pairwise */ 2574827bd09bSSatish Balay if (gs->num_pairs) 2575827bd09bSSatish Balay {gs_gop_pairwise_times(gs,vals);} 2576827bd09bSSatish Balay 2577827bd09bSSatish Balay /* tree */ 2578827bd09bSSatish Balay else if (gs->max_left_over) 2579827bd09bSSatish Balay {gs_gop_tree_times(gs,vals);} 2580827bd09bSSatish Balay } 25813fdc5746SBarry Smith PetscFunctionReturn(0); 2582827bd09bSSatish Balay } 2583827bd09bSSatish Balay 2584827bd09bSSatish Balay 2585827bd09bSSatish Balay 2586827bd09bSSatish Balay /****************************************************************************** 2587827bd09bSSatish Balay Function: gather_scatter 2588827bd09bSSatish Balay 2589827bd09bSSatish Balay Input : 2590827bd09bSSatish Balay Output: 2591827bd09bSSatish Balay Return: 2592827bd09bSSatish Balay Description: 2593827bd09bSSatish Balay ******************************************************************************/ 25940924e98cSBarry Smith static PetscErrorCode gs_gop_local_times( gs_id *gs, PetscScalar *vals) 2595827bd09bSSatish Balay { 2596a501084fSBarry Smith int *num, *map, **reduce; 2597a501084fSBarry Smith PetscScalar tmp; 2598827bd09bSSatish Balay 25993fdc5746SBarry Smith PetscFunctionBegin; 2600827bd09bSSatish Balay num = gs->num_local_reduce; 2601827bd09bSSatish Balay reduce = gs->local_reduce; 2602827bd09bSSatish Balay while ((map = *reduce)) 2603827bd09bSSatish Balay { 2604827bd09bSSatish Balay /* wall */ 2605827bd09bSSatish Balay if (*num == 2) 2606827bd09bSSatish Balay { 2607827bd09bSSatish Balay num ++; reduce++; 2608827bd09bSSatish Balay vals[map[1]] = vals[map[0]] *= vals[map[1]]; 2609827bd09bSSatish Balay } 2610827bd09bSSatish Balay /* corner shared by three elements */ 2611827bd09bSSatish Balay else if (*num == 3) 2612827bd09bSSatish Balay { 2613827bd09bSSatish Balay num ++; reduce++; 2614827bd09bSSatish Balay vals[map[2]]=vals[map[1]]=vals[map[0]]*=(vals[map[1]]*vals[map[2]]); 2615827bd09bSSatish Balay } 2616827bd09bSSatish Balay /* corner shared by four elements */ 2617827bd09bSSatish Balay else if (*num == 4) 2618827bd09bSSatish Balay { 2619827bd09bSSatish Balay num ++; reduce++; 2620827bd09bSSatish Balay vals[map[1]]=vals[map[2]]=vals[map[3]]=vals[map[0]] *= 2621827bd09bSSatish Balay (vals[map[1]] * vals[map[2]] * vals[map[3]]); 2622827bd09bSSatish Balay } 2623827bd09bSSatish Balay /* general case ... odd geoms ... 3D*/ 2624827bd09bSSatish Balay else 2625827bd09bSSatish Balay { 2626827bd09bSSatish Balay num ++; 2627827bd09bSSatish Balay tmp = 1.0; 2628827bd09bSSatish Balay while (*map >= 0) 2629827bd09bSSatish Balay {tmp *= *(vals + *map++);} 2630827bd09bSSatish Balay 2631827bd09bSSatish Balay map = *reduce++; 2632827bd09bSSatish Balay while (*map >= 0) 2633827bd09bSSatish Balay {*(vals + *map++) = tmp;} 2634827bd09bSSatish Balay } 2635827bd09bSSatish Balay } 26363fdc5746SBarry Smith PetscFunctionReturn(0); 2637827bd09bSSatish Balay } 2638827bd09bSSatish Balay 2639827bd09bSSatish Balay 2640827bd09bSSatish Balay 2641827bd09bSSatish Balay /****************************************************************************** 2642827bd09bSSatish Balay Function: gather_scatter 2643827bd09bSSatish Balay 2644827bd09bSSatish Balay Input : 2645827bd09bSSatish Balay Output: 2646827bd09bSSatish Balay Return: 2647827bd09bSSatish Balay Description: 2648827bd09bSSatish Balay ******************************************************************************/ 26490924e98cSBarry Smith static PetscErrorCode gs_gop_local_in_times( gs_id *gs, PetscScalar *vals) 2650827bd09bSSatish Balay { 2651a501084fSBarry Smith int *num, *map, **reduce; 2652a501084fSBarry Smith PetscScalar *base; 2653827bd09bSSatish Balay 26543fdc5746SBarry Smith PetscFunctionBegin; 2655827bd09bSSatish Balay num = gs->num_gop_local_reduce; 2656827bd09bSSatish Balay reduce = gs->gop_local_reduce; 2657827bd09bSSatish Balay while ((map = *reduce++)) 2658827bd09bSSatish Balay { 2659827bd09bSSatish Balay /* wall */ 2660827bd09bSSatish Balay if (*num == 2) 2661827bd09bSSatish Balay { 2662827bd09bSSatish Balay num ++; 2663827bd09bSSatish Balay vals[map[0]] *= vals[map[1]]; 2664827bd09bSSatish Balay } 2665827bd09bSSatish Balay /* corner shared by three elements */ 2666827bd09bSSatish Balay else if (*num == 3) 2667827bd09bSSatish Balay { 2668827bd09bSSatish Balay num ++; 2669827bd09bSSatish Balay vals[map[0]] *= (vals[map[1]] * vals[map[2]]); 2670827bd09bSSatish Balay } 2671827bd09bSSatish Balay /* corner shared by four elements */ 2672827bd09bSSatish Balay else if (*num == 4) 2673827bd09bSSatish Balay { 2674827bd09bSSatish Balay num ++; 2675827bd09bSSatish Balay vals[map[0]] *= (vals[map[1]] * vals[map[2]] * vals[map[3]]); 2676827bd09bSSatish Balay } 2677827bd09bSSatish Balay /* general case ... odd geoms ... 3D*/ 2678827bd09bSSatish Balay else 2679827bd09bSSatish Balay { 2680827bd09bSSatish Balay num++; 2681827bd09bSSatish Balay base = vals + *map++; 2682827bd09bSSatish Balay while (*map >= 0) 2683827bd09bSSatish Balay {*base *= *(vals + *map++);} 2684827bd09bSSatish Balay } 2685827bd09bSSatish Balay } 26863fdc5746SBarry Smith PetscFunctionReturn(0); 2687827bd09bSSatish Balay } 2688827bd09bSSatish Balay 2689827bd09bSSatish Balay 2690827bd09bSSatish Balay 2691827bd09bSSatish Balay /****************************************************************************** 2692827bd09bSSatish Balay Function: gather_scatter 2693827bd09bSSatish Balay 2694827bd09bSSatish Balay VERSION 3 :: 2695827bd09bSSatish Balay 2696827bd09bSSatish Balay Input : 2697827bd09bSSatish Balay Output: 2698827bd09bSSatish Balay Return: 2699827bd09bSSatish Balay Description: 2700827bd09bSSatish Balay ******************************************************************************/ 27010924e98cSBarry Smith static PetscErrorCode gs_gop_pairwise_times( gs_id *gs, PetscScalar *in_vals) 2702827bd09bSSatish Balay { 2703a501084fSBarry Smith PetscScalar *dptr1, *dptr2, *dptr3, *in1, *in2; 2704a501084fSBarry Smith int *iptr, *msg_list, *msg_size, **msg_nodes; 2705a501084fSBarry Smith int *pw, *list, *size, **nodes; 2706827bd09bSSatish Balay MPI_Request *msg_ids_in, *msg_ids_out, *ids_in, *ids_out; 2707827bd09bSSatish Balay MPI_Status status; 27083fdc5746SBarry Smith PetscErrorCode ierr; 2709827bd09bSSatish Balay 27103fdc5746SBarry Smith PetscFunctionBegin; 2711a501084fSBarry Smith /* strip and load s */ 2712827bd09bSSatish Balay msg_list =list = gs->pair_list; 2713827bd09bSSatish Balay msg_size =size = gs->msg_sizes; 2714827bd09bSSatish Balay msg_nodes=nodes = gs->node_list; 2715827bd09bSSatish Balay iptr=pw = gs->pw_elm_list; 2716827bd09bSSatish Balay dptr1=dptr3 = gs->pw_vals; 2717827bd09bSSatish Balay msg_ids_in = ids_in = gs->msg_ids_in; 2718827bd09bSSatish Balay msg_ids_out = ids_out = gs->msg_ids_out; 2719827bd09bSSatish Balay dptr2 = gs->out; 2720827bd09bSSatish Balay in1=in2 = gs->in; 2721827bd09bSSatish Balay 2722827bd09bSSatish Balay /* post the receives */ 2723827bd09bSSatish Balay /* msg_nodes=nodes; */ 2724827bd09bSSatish Balay do 2725827bd09bSSatish Balay { 2726827bd09bSSatish Balay /* Should MPI_ANY_SOURCE be replaced by *list ? In that case do the 2727827bd09bSSatish Balay second one *list and do list++ afterwards */ 27283fdc5746SBarry Smith ierr = MPI_Irecv(in1, *size, MPIU_SCALAR, MPI_ANY_SOURCE, MSGTAG1 + *list++, gs->gs_comm, msg_ids_in++);CHKERRQ(ierr); 2729827bd09bSSatish Balay in1 += *size++; 2730827bd09bSSatish Balay } 2731827bd09bSSatish Balay while (*++msg_nodes); 2732827bd09bSSatish Balay msg_nodes=nodes; 2733827bd09bSSatish Balay 2734827bd09bSSatish Balay /* load gs values into in out gs buffers */ 2735827bd09bSSatish Balay while (*iptr >= 0) 2736827bd09bSSatish Balay {*dptr3++ = *(in_vals + *iptr++);} 2737827bd09bSSatish Balay 2738827bd09bSSatish Balay /* load out buffers and post the sends */ 2739827bd09bSSatish Balay while ((iptr = *msg_nodes++)) 2740827bd09bSSatish Balay { 2741827bd09bSSatish Balay dptr3 = dptr2; 2742827bd09bSSatish Balay while (*iptr >= 0) 2743827bd09bSSatish Balay {*dptr2++ = *(dptr1 + *iptr++);} 2744827bd09bSSatish Balay /* CHECK PERSISTENT COMMS MODE FOR ALL THIS STUFF */ 2745827bd09bSSatish Balay /* is msg_ids_out++ correct? */ 27463fdc5746SBarry Smith ierr = MPI_Isend(dptr3, *msg_size++, MPIU_SCALAR, *msg_list++, MSGTAG1+my_id, gs->gs_comm, msg_ids_out++);CHKERRQ(ierr); 2747827bd09bSSatish Balay } 2748827bd09bSSatish Balay 2749827bd09bSSatish Balay if (gs->max_left_over) 2750827bd09bSSatish Balay {gs_gop_tree_times(gs,in_vals);} 2751827bd09bSSatish Balay 2752827bd09bSSatish Balay /* process the received data */ 2753827bd09bSSatish Balay msg_nodes=nodes; 2754827bd09bSSatish Balay while ((iptr = *nodes++)) 2755827bd09bSSatish Balay { 2756827bd09bSSatish Balay /* Should I check the return value of MPI_Wait() or status? */ 2757827bd09bSSatish Balay /* Can this loop be replaced by a call to MPI_Waitall()? */ 27583fdc5746SBarry Smith ierr = MPI_Wait(ids_in++, &status);CHKERRQ(ierr); 2759827bd09bSSatish Balay while (*iptr >= 0) 2760827bd09bSSatish Balay {*(dptr1 + *iptr++) *= *in2++;} 2761827bd09bSSatish Balay } 2762827bd09bSSatish Balay 2763827bd09bSSatish Balay /* replace vals */ 2764827bd09bSSatish Balay while (*pw >= 0) 2765827bd09bSSatish Balay {*(in_vals + *pw++) = *dptr1++;} 2766827bd09bSSatish Balay 2767827bd09bSSatish Balay /* clear isend message handles */ 2768827bd09bSSatish Balay /* This changed for clarity though it could be the same */ 2769827bd09bSSatish Balay while (*msg_nodes++) 2770827bd09bSSatish Balay /* Should I check the return value of MPI_Wait() or status? */ 2771827bd09bSSatish Balay /* Can this loop be replaced by a call to MPI_Waitall()? */ 27723fdc5746SBarry Smith {ierr = MPI_Wait(ids_out++, &status);CHKERRQ(ierr);} 27733fdc5746SBarry Smith PetscFunctionReturn(0); 2774827bd09bSSatish Balay } 2775827bd09bSSatish Balay 2776827bd09bSSatish Balay 2777827bd09bSSatish Balay 2778827bd09bSSatish Balay /****************************************************************************** 2779827bd09bSSatish Balay Function: gather_scatter 2780827bd09bSSatish Balay 2781827bd09bSSatish Balay Input : 2782827bd09bSSatish Balay Output: 2783827bd09bSSatish Balay Return: 2784827bd09bSSatish Balay Description: 2785827bd09bSSatish Balay ******************************************************************************/ 27860924e98cSBarry Smith static PetscErrorCode gs_gop_tree_times(gs_id *gs, PetscScalar *vals) 2787827bd09bSSatish Balay { 2788827bd09bSSatish Balay int size; 2789827bd09bSSatish Balay int *in, *out; 2790a501084fSBarry Smith PetscScalar *buf, *work; 27913fdc5746SBarry Smith PetscErrorCode ierr; 2792827bd09bSSatish Balay 27933fdc5746SBarry Smith PetscFunctionBegin; 2794827bd09bSSatish Balay in = gs->tree_map_in; 2795827bd09bSSatish Balay out = gs->tree_map_out; 2796827bd09bSSatish Balay buf = gs->tree_buf; 2797827bd09bSSatish Balay work = gs->tree_work; 2798827bd09bSSatish Balay size = gs->tree_nel; 2799827bd09bSSatish Balay 2800827bd09bSSatish Balay rvec_one(buf,size); 2801827bd09bSSatish Balay 2802827bd09bSSatish Balay while (*in >= 0) 2803827bd09bSSatish Balay {*(buf + *out++) = *(vals + *in++);} 2804827bd09bSSatish Balay 2805827bd09bSSatish Balay in = gs->tree_map_in; 2806827bd09bSSatish Balay out = gs->tree_map_out; 28073fdc5746SBarry Smith ierr = MPI_Allreduce(buf,work,size,MPIU_SCALAR,MPI_PROD,gs->gs_comm);CHKERRQ(ierr); 2808827bd09bSSatish Balay while (*in >= 0) 2809827bd09bSSatish Balay {*(vals + *in++) = *(work + *out++);} 28103fdc5746SBarry Smith PetscFunctionReturn(0); 2811827bd09bSSatish Balay } 2812827bd09bSSatish Balay 2813827bd09bSSatish Balay 2814827bd09bSSatish Balay 2815827bd09bSSatish Balay /****************************************************************************** 2816827bd09bSSatish Balay Function: gather_scatter 2817827bd09bSSatish Balay 2818827bd09bSSatish Balay 2819827bd09bSSatish Balay Input : 2820827bd09bSSatish Balay Output: 2821827bd09bSSatish Balay Return: 2822827bd09bSSatish Balay Description: 2823827bd09bSSatish Balay ******************************************************************************/ 28240924e98cSBarry Smith static PetscErrorCode gs_gop_plus( gs_id *gs, PetscScalar *vals) 2825827bd09bSSatish Balay { 28263fdc5746SBarry Smith PetscFunctionBegin; 2827827bd09bSSatish Balay /* local only operations!!! */ 2828827bd09bSSatish Balay if (gs->num_local) 2829827bd09bSSatish Balay {gs_gop_local_plus(gs,vals);} 2830827bd09bSSatish Balay 2831827bd09bSSatish Balay /* if intersection tree/pairwise and local isn't empty */ 2832827bd09bSSatish Balay if (gs->num_local_gop) 2833827bd09bSSatish Balay { 2834827bd09bSSatish Balay gs_gop_local_in_plus(gs,vals); 2835827bd09bSSatish Balay 2836827bd09bSSatish Balay /* pairwise will NOT do tree inside ... */ 2837827bd09bSSatish Balay if (gs->num_pairs) 2838827bd09bSSatish Balay {gs_gop_pairwise_plus(gs,vals);} 2839827bd09bSSatish Balay 2840827bd09bSSatish Balay /* tree */ 2841827bd09bSSatish Balay if (gs->max_left_over) 2842827bd09bSSatish Balay {gs_gop_tree_plus(gs,vals);} 2843827bd09bSSatish Balay 2844827bd09bSSatish Balay gs_gop_local_out(gs,vals); 2845827bd09bSSatish Balay } 2846827bd09bSSatish Balay /* if intersection tree/pairwise and local is empty */ 2847827bd09bSSatish Balay else 2848827bd09bSSatish Balay { 2849827bd09bSSatish Balay /* pairwise will NOT do tree inside */ 2850827bd09bSSatish Balay if (gs->num_pairs) 2851827bd09bSSatish Balay {gs_gop_pairwise_plus(gs,vals);} 2852827bd09bSSatish Balay 2853827bd09bSSatish Balay /* tree */ 2854827bd09bSSatish Balay if (gs->max_left_over) 2855827bd09bSSatish Balay {gs_gop_tree_plus(gs,vals);} 2856827bd09bSSatish Balay } 28573fdc5746SBarry Smith PetscFunctionReturn(0); 2858827bd09bSSatish Balay } 2859827bd09bSSatish Balay 2860827bd09bSSatish Balay 2861827bd09bSSatish Balay 2862827bd09bSSatish Balay /****************************************************************************** 2863827bd09bSSatish Balay Function: gather_scatter 2864827bd09bSSatish Balay 2865827bd09bSSatish Balay Input : 2866827bd09bSSatish Balay Output: 2867827bd09bSSatish Balay Return: 2868827bd09bSSatish Balay Description: 2869827bd09bSSatish Balay ******************************************************************************/ 28700924e98cSBarry Smith static PetscErrorCode gs_gop_local_plus( gs_id *gs, PetscScalar *vals) 2871827bd09bSSatish Balay { 2872a501084fSBarry Smith int *num, *map, **reduce; 2873a501084fSBarry Smith PetscScalar tmp; 2874827bd09bSSatish Balay 28753fdc5746SBarry Smith PetscFunctionBegin; 2876827bd09bSSatish Balay num = gs->num_local_reduce; 2877827bd09bSSatish Balay reduce = gs->local_reduce; 2878827bd09bSSatish Balay while ((map = *reduce)) 2879827bd09bSSatish Balay { 2880827bd09bSSatish Balay /* wall */ 2881827bd09bSSatish Balay if (*num == 2) 2882827bd09bSSatish Balay { 2883827bd09bSSatish Balay num ++; reduce++; 2884827bd09bSSatish Balay vals[map[1]] = vals[map[0]] += vals[map[1]]; 2885827bd09bSSatish Balay } 2886827bd09bSSatish Balay /* corner shared by three elements */ 2887827bd09bSSatish Balay else if (*num == 3) 2888827bd09bSSatish Balay { 2889827bd09bSSatish Balay num ++; reduce++; 2890827bd09bSSatish Balay vals[map[2]]=vals[map[1]]=vals[map[0]]+=(vals[map[1]]+vals[map[2]]); 2891827bd09bSSatish Balay } 2892827bd09bSSatish Balay /* corner shared by four elements */ 2893827bd09bSSatish Balay else if (*num == 4) 2894827bd09bSSatish Balay { 2895827bd09bSSatish Balay num ++; reduce++; 2896827bd09bSSatish Balay vals[map[1]]=vals[map[2]]=vals[map[3]]=vals[map[0]] += 2897827bd09bSSatish Balay (vals[map[1]] + vals[map[2]] + vals[map[3]]); 2898827bd09bSSatish Balay } 2899827bd09bSSatish Balay /* general case ... odd geoms ... 3D*/ 2900827bd09bSSatish Balay else 2901827bd09bSSatish Balay { 2902827bd09bSSatish Balay num ++; 2903827bd09bSSatish Balay tmp = 0.0; 2904827bd09bSSatish Balay while (*map >= 0) 2905827bd09bSSatish Balay {tmp += *(vals + *map++);} 2906827bd09bSSatish Balay 2907827bd09bSSatish Balay map = *reduce++; 2908827bd09bSSatish Balay while (*map >= 0) 2909827bd09bSSatish Balay {*(vals + *map++) = tmp;} 2910827bd09bSSatish Balay } 2911827bd09bSSatish Balay } 29123fdc5746SBarry Smith PetscFunctionReturn(0); 2913827bd09bSSatish Balay } 2914827bd09bSSatish Balay 2915827bd09bSSatish Balay 2916827bd09bSSatish Balay 2917827bd09bSSatish Balay /****************************************************************************** 2918827bd09bSSatish Balay Function: gather_scatter 2919827bd09bSSatish Balay 2920827bd09bSSatish Balay Input : 2921827bd09bSSatish Balay Output: 2922827bd09bSSatish Balay Return: 2923827bd09bSSatish Balay Description: 2924827bd09bSSatish Balay ******************************************************************************/ 29250924e98cSBarry Smith static PetscErrorCode gs_gop_local_in_plus( gs_id *gs, PetscScalar *vals) 2926827bd09bSSatish Balay { 2927a501084fSBarry Smith int *num, *map, **reduce; 2928a501084fSBarry Smith PetscScalar *base; 2929827bd09bSSatish Balay 29303fdc5746SBarry Smith PetscFunctionBegin; 2931827bd09bSSatish Balay num = gs->num_gop_local_reduce; 2932827bd09bSSatish Balay reduce = gs->gop_local_reduce; 2933827bd09bSSatish Balay while ((map = *reduce++)) 2934827bd09bSSatish Balay { 2935827bd09bSSatish Balay /* wall */ 2936827bd09bSSatish Balay if (*num == 2) 2937827bd09bSSatish Balay { 2938827bd09bSSatish Balay num ++; 2939827bd09bSSatish Balay vals[map[0]] += vals[map[1]]; 2940827bd09bSSatish Balay } 2941827bd09bSSatish Balay /* corner shared by three elements */ 2942827bd09bSSatish Balay else if (*num == 3) 2943827bd09bSSatish Balay { 2944827bd09bSSatish Balay num ++; 2945827bd09bSSatish Balay vals[map[0]] += (vals[map[1]] + vals[map[2]]); 2946827bd09bSSatish Balay } 2947827bd09bSSatish Balay /* corner shared by four elements */ 2948827bd09bSSatish Balay else if (*num == 4) 2949827bd09bSSatish Balay { 2950827bd09bSSatish Balay num ++; 2951827bd09bSSatish Balay vals[map[0]] += (vals[map[1]] + vals[map[2]] + vals[map[3]]); 2952827bd09bSSatish Balay } 2953827bd09bSSatish Balay /* general case ... odd geoms ... 3D*/ 2954827bd09bSSatish Balay else 2955827bd09bSSatish Balay { 2956827bd09bSSatish Balay num++; 2957827bd09bSSatish Balay base = vals + *map++; 2958827bd09bSSatish Balay while (*map >= 0) 2959827bd09bSSatish Balay {*base += *(vals + *map++);} 2960827bd09bSSatish Balay } 2961827bd09bSSatish Balay } 29623fdc5746SBarry Smith PetscFunctionReturn(0); 2963827bd09bSSatish Balay } 2964827bd09bSSatish Balay 2965827bd09bSSatish Balay 2966827bd09bSSatish Balay 2967827bd09bSSatish Balay /****************************************************************************** 2968827bd09bSSatish Balay Function: gather_scatter 2969827bd09bSSatish Balay 2970827bd09bSSatish Balay VERSION 3 :: 2971827bd09bSSatish Balay 2972827bd09bSSatish Balay Input : 2973827bd09bSSatish Balay Output: 2974827bd09bSSatish Balay Return: 2975827bd09bSSatish Balay Description: 2976827bd09bSSatish Balay ******************************************************************************/ 29770924e98cSBarry Smith static PetscErrorCode gs_gop_pairwise_plus( gs_id *gs, PetscScalar *in_vals) 2978827bd09bSSatish Balay { 2979a501084fSBarry Smith PetscScalar *dptr1, *dptr2, *dptr3, *in1, *in2; 2980a501084fSBarry Smith int *iptr, *msg_list, *msg_size, **msg_nodes; 2981a501084fSBarry Smith int *pw, *list, *size, **nodes; 2982827bd09bSSatish Balay MPI_Request *msg_ids_in, *msg_ids_out, *ids_in, *ids_out; 2983827bd09bSSatish Balay MPI_Status status; 29843fdc5746SBarry Smith PetscErrorCode ierr; 2985827bd09bSSatish Balay 29863fdc5746SBarry Smith PetscFunctionBegin; 2987a501084fSBarry Smith /* strip and load s */ 2988827bd09bSSatish Balay msg_list =list = gs->pair_list; 2989827bd09bSSatish Balay msg_size =size = gs->msg_sizes; 2990827bd09bSSatish Balay msg_nodes=nodes = gs->node_list; 2991827bd09bSSatish Balay iptr=pw = gs->pw_elm_list; 2992827bd09bSSatish Balay dptr1=dptr3 = gs->pw_vals; 2993827bd09bSSatish Balay msg_ids_in = ids_in = gs->msg_ids_in; 2994827bd09bSSatish Balay msg_ids_out = ids_out = gs->msg_ids_out; 2995827bd09bSSatish Balay dptr2 = gs->out; 2996827bd09bSSatish Balay in1=in2 = gs->in; 2997827bd09bSSatish Balay 2998827bd09bSSatish Balay /* post the receives */ 2999827bd09bSSatish Balay /* msg_nodes=nodes; */ 3000827bd09bSSatish Balay do 3001827bd09bSSatish Balay { 3002827bd09bSSatish Balay /* Should MPI_ANY_SOURCE be replaced by *list ? In that case do the 3003827bd09bSSatish Balay second one *list and do list++ afterwards */ 30043fdc5746SBarry Smith ierr = MPI_Irecv(in1, *size, MPIU_SCALAR, MPI_ANY_SOURCE, MSGTAG1 + *list++, gs->gs_comm, msg_ids_in++);CHKERRQ(ierr); 3005827bd09bSSatish Balay in1 += *size++; 3006827bd09bSSatish Balay } 3007827bd09bSSatish Balay while (*++msg_nodes); 3008827bd09bSSatish Balay msg_nodes=nodes; 3009827bd09bSSatish Balay 3010827bd09bSSatish Balay /* load gs values into in out gs buffers */ 3011827bd09bSSatish Balay while (*iptr >= 0) 3012827bd09bSSatish Balay {*dptr3++ = *(in_vals + *iptr++);} 3013827bd09bSSatish Balay 3014827bd09bSSatish Balay /* load out buffers and post the sends */ 3015827bd09bSSatish Balay while ((iptr = *msg_nodes++)) 3016827bd09bSSatish Balay { 3017827bd09bSSatish Balay dptr3 = dptr2; 3018827bd09bSSatish Balay while (*iptr >= 0) 3019827bd09bSSatish Balay {*dptr2++ = *(dptr1 + *iptr++);} 3020827bd09bSSatish Balay /* CHECK PERSISTENT COMMS MODE FOR ALL THIS STUFF */ 3021827bd09bSSatish Balay /* is msg_ids_out++ correct? */ 30223fdc5746SBarry Smith ierr = MPI_Isend(dptr3, *msg_size++, MPIU_SCALAR, *msg_list++, MSGTAG1+my_id, gs->gs_comm, msg_ids_out++);CHKERRQ(ierr); 3023827bd09bSSatish Balay } 3024827bd09bSSatish Balay 3025827bd09bSSatish Balay /* do the tree while we're waiting */ 3026827bd09bSSatish Balay if (gs->max_left_over) 3027827bd09bSSatish Balay {gs_gop_tree_plus(gs,in_vals);} 3028827bd09bSSatish Balay 3029827bd09bSSatish Balay /* process the received data */ 3030827bd09bSSatish Balay msg_nodes=nodes; 3031827bd09bSSatish Balay while ((iptr = *nodes++)) 3032827bd09bSSatish Balay { 3033827bd09bSSatish Balay /* Should I check the return value of MPI_Wait() or status? */ 3034827bd09bSSatish Balay /* Can this loop be replaced by a call to MPI_Waitall()? */ 30353fdc5746SBarry Smith ierr = MPI_Wait(ids_in++, &status);CHKERRQ(ierr); 3036827bd09bSSatish Balay while (*iptr >= 0) 3037827bd09bSSatish Balay {*(dptr1 + *iptr++) += *in2++;} 3038827bd09bSSatish Balay } 3039827bd09bSSatish Balay 3040827bd09bSSatish Balay /* replace vals */ 3041827bd09bSSatish Balay while (*pw >= 0) 3042827bd09bSSatish Balay {*(in_vals + *pw++) = *dptr1++;} 3043827bd09bSSatish Balay 3044827bd09bSSatish Balay /* clear isend message handles */ 3045827bd09bSSatish Balay /* This changed for clarity though it could be the same */ 3046827bd09bSSatish Balay while (*msg_nodes++) 3047827bd09bSSatish Balay /* Should I check the return value of MPI_Wait() or status? */ 3048827bd09bSSatish Balay /* Can this loop be replaced by a call to MPI_Waitall()? */ 30493fdc5746SBarry Smith {ierr = MPI_Wait(ids_out++, &status);CHKERRQ(ierr);} 30503fdc5746SBarry Smith PetscFunctionReturn(0); 3051827bd09bSSatish Balay } 3052827bd09bSSatish Balay 3053827bd09bSSatish Balay 3054827bd09bSSatish Balay 3055827bd09bSSatish Balay /****************************************************************************** 3056827bd09bSSatish Balay Function: gather_scatter 3057827bd09bSSatish Balay 3058827bd09bSSatish Balay Input : 3059827bd09bSSatish Balay Output: 3060827bd09bSSatish Balay Return: 3061827bd09bSSatish Balay Description: 3062827bd09bSSatish Balay ******************************************************************************/ 30630924e98cSBarry Smith static PetscErrorCode gs_gop_tree_plus(gs_id *gs, PetscScalar *vals) 3064827bd09bSSatish Balay { 3065827bd09bSSatish Balay int size; 3066827bd09bSSatish Balay int *in, *out; 3067a501084fSBarry Smith PetscScalar *buf, *work; 30683fdc5746SBarry Smith PetscErrorCode ierr; 3069827bd09bSSatish Balay 30703fdc5746SBarry Smith PetscFunctionBegin; 3071827bd09bSSatish Balay in = gs->tree_map_in; 3072827bd09bSSatish Balay out = gs->tree_map_out; 3073827bd09bSSatish Balay buf = gs->tree_buf; 3074827bd09bSSatish Balay work = gs->tree_work; 3075827bd09bSSatish Balay size = gs->tree_nel; 3076827bd09bSSatish Balay 3077827bd09bSSatish Balay rvec_zero(buf,size); 3078827bd09bSSatish Balay 3079827bd09bSSatish Balay while (*in >= 0) 3080827bd09bSSatish Balay {*(buf + *out++) = *(vals + *in++);} 3081827bd09bSSatish Balay 3082827bd09bSSatish Balay in = gs->tree_map_in; 3083827bd09bSSatish Balay out = gs->tree_map_out; 30843fdc5746SBarry Smith ierr = MPI_Allreduce(buf,work,size,MPIU_SCALAR,MPI_SUM,gs->gs_comm);CHKERRQ(ierr); 3085827bd09bSSatish Balay while (*in >= 0) 3086827bd09bSSatish Balay {*(vals + *in++) = *(work + *out++);} 30873fdc5746SBarry Smith PetscFunctionReturn(0); 3088827bd09bSSatish Balay } 3089827bd09bSSatish Balay 3090827bd09bSSatish Balay /****************************************************************************** 3091827bd09bSSatish Balay Function: gs_free() 3092827bd09bSSatish Balay 3093827bd09bSSatish Balay Input : 3094827bd09bSSatish Balay 3095827bd09bSSatish Balay Output: 3096827bd09bSSatish Balay 3097827bd09bSSatish Balay Return: 3098827bd09bSSatish Balay 3099827bd09bSSatish Balay Description: 3100a501084fSBarry Smith if (gs->sss) {free((void*) gs->sss);} 3101827bd09bSSatish Balay ******************************************************************************/ 31020924e98cSBarry Smith PetscErrorCode gs_free( gs_id *gs) 3103827bd09bSSatish Balay { 3104a501084fSBarry Smith int i; 3105827bd09bSSatish Balay 31063fdc5746SBarry Smith PetscFunctionBegin; 3107a501084fSBarry Smith if (gs->nghs) {free((void*) gs->nghs);} 3108a501084fSBarry Smith if (gs->pw_nghs) {free((void*) gs->pw_nghs);} 3109827bd09bSSatish Balay 3110827bd09bSSatish Balay /* tree */ 3111827bd09bSSatish Balay if (gs->max_left_over) 3112827bd09bSSatish Balay { 3113a501084fSBarry Smith if (gs->tree_elms) {free((void*) gs->tree_elms);} 3114a501084fSBarry Smith if (gs->tree_buf) {free((void*) gs->tree_buf);} 3115a501084fSBarry Smith if (gs->tree_work) {free((void*) gs->tree_work);} 3116a501084fSBarry Smith if (gs->tree_map_in) {free((void*) gs->tree_map_in);} 3117a501084fSBarry Smith if (gs->tree_map_out) {free((void*) gs->tree_map_out);} 3118827bd09bSSatish Balay } 3119827bd09bSSatish Balay 3120827bd09bSSatish Balay /* pairwise info */ 3121827bd09bSSatish Balay if (gs->num_pairs) 3122827bd09bSSatish Balay { 3123827bd09bSSatish Balay /* should be NULL already */ 3124a501084fSBarry Smith if (gs->ngh_buf) {free((void*) gs->ngh_buf);} 3125a501084fSBarry Smith if (gs->elms) {free((void*) gs->elms);} 3126a501084fSBarry Smith if (gs->local_elms) {free((void*) gs->local_elms);} 3127a501084fSBarry Smith if (gs->companion) {free((void*) gs->companion);} 3128827bd09bSSatish Balay 3129827bd09bSSatish Balay /* only set if pairwise */ 3130a501084fSBarry Smith if (gs->vals) {free((void*) gs->vals);} 3131a501084fSBarry Smith if (gs->in) {free((void*) gs->in);} 3132a501084fSBarry Smith if (gs->out) {free((void*) gs->out);} 3133a501084fSBarry Smith if (gs->msg_ids_in) {free((void*) gs->msg_ids_in);} 3134a501084fSBarry Smith if (gs->msg_ids_out) {free((void*) gs->msg_ids_out);} 3135a501084fSBarry Smith if (gs->pw_vals) {free((void*) gs->pw_vals);} 3136a501084fSBarry Smith if (gs->pw_elm_list) {free((void*) gs->pw_elm_list);} 3137827bd09bSSatish Balay if (gs->node_list) 3138827bd09bSSatish Balay { 3139827bd09bSSatish Balay for (i=0;i<gs->num_pairs;i++) 3140a501084fSBarry Smith {if (gs->node_list[i]) {free((void*) gs->node_list[i]);}} 3141a501084fSBarry Smith free((void*) gs->node_list); 3142827bd09bSSatish Balay } 3143a501084fSBarry Smith if (gs->msg_sizes) {free((void*) gs->msg_sizes);} 3144a501084fSBarry Smith if (gs->pair_list) {free((void*) gs->pair_list);} 3145827bd09bSSatish Balay } 3146827bd09bSSatish Balay 3147827bd09bSSatish Balay /* local info */ 3148827bd09bSSatish Balay if (gs->num_local_total>=0) 3149827bd09bSSatish Balay { 3150827bd09bSSatish Balay for (i=0;i<gs->num_local_total+1;i++) 3151827bd09bSSatish Balay /* for (i=0;i<gs->num_local_total;i++) */ 3152827bd09bSSatish Balay { 3153827bd09bSSatish Balay if (gs->num_gop_local_reduce[i]) 3154a501084fSBarry Smith {free((void*) gs->gop_local_reduce[i]);} 3155827bd09bSSatish Balay } 3156827bd09bSSatish Balay } 3157827bd09bSSatish Balay 3158827bd09bSSatish Balay /* if intersection tree/pairwise and local isn't empty */ 3159a501084fSBarry Smith if (gs->gop_local_reduce) {free((void*) gs->gop_local_reduce);} 3160a501084fSBarry Smith if (gs->num_gop_local_reduce) {free((void*) gs->num_gop_local_reduce);} 3161827bd09bSSatish Balay 3162a501084fSBarry Smith free((void*) gs); 31633fdc5746SBarry Smith PetscFunctionReturn(0); 3164827bd09bSSatish Balay } 3165827bd09bSSatish Balay 3166827bd09bSSatish Balay 3167827bd09bSSatish Balay 3168827bd09bSSatish Balay 3169827bd09bSSatish Balay 3170827bd09bSSatish Balay 3171827bd09bSSatish Balay /****************************************************************************** 3172827bd09bSSatish Balay Function: gather_scatter 3173827bd09bSSatish Balay 3174827bd09bSSatish Balay Input : 3175827bd09bSSatish Balay Output: 3176827bd09bSSatish Balay Return: 3177827bd09bSSatish Balay Description: 3178827bd09bSSatish Balay ******************************************************************************/ 31790924e98cSBarry Smith PetscErrorCode gs_gop_vec( gs_id *gs, PetscScalar *vals, const char *op, int step) 3180827bd09bSSatish Balay { 31813fdc5746SBarry Smith PetscFunctionBegin; 3182827bd09bSSatish Balay switch (*op) { 3183827bd09bSSatish Balay case '+': 3184827bd09bSSatish Balay gs_gop_vec_plus(gs,vals,step); 3185827bd09bSSatish Balay break; 3186827bd09bSSatish Balay default: 3187827bd09bSSatish Balay error_msg_warning("gs_gop_vec() :: %c is not a valid op",op[0]); 3188827bd09bSSatish Balay error_msg_warning("gs_gop_vec() :: default :: plus"); 3189827bd09bSSatish Balay gs_gop_vec_plus(gs,vals,step); 3190827bd09bSSatish Balay break; 3191827bd09bSSatish Balay } 31923fdc5746SBarry Smith PetscFunctionReturn(0); 3193827bd09bSSatish Balay } 3194827bd09bSSatish Balay 3195827bd09bSSatish Balay 3196827bd09bSSatish Balay 3197827bd09bSSatish Balay /****************************************************************************** 3198827bd09bSSatish Balay Function: gather_scatter 3199827bd09bSSatish Balay 3200827bd09bSSatish Balay Input : 3201827bd09bSSatish Balay Output: 3202827bd09bSSatish Balay Return: 3203827bd09bSSatish Balay Description: 3204827bd09bSSatish Balay ******************************************************************************/ 32050924e98cSBarry Smith static PetscErrorCode gs_gop_vec_plus( gs_id *gs, PetscScalar *vals, int step) 3206827bd09bSSatish Balay { 32073fdc5746SBarry Smith PetscFunctionBegin; 3208827bd09bSSatish Balay if (!gs) {error_msg_fatal("gs_gop_vec() passed NULL gs handle!!!");} 3209827bd09bSSatish Balay 3210827bd09bSSatish Balay /* local only operations!!! */ 3211827bd09bSSatish Balay if (gs->num_local) 3212827bd09bSSatish Balay {gs_gop_vec_local_plus(gs,vals,step);} 3213827bd09bSSatish Balay 3214827bd09bSSatish Balay /* if intersection tree/pairwise and local isn't empty */ 3215827bd09bSSatish Balay if (gs->num_local_gop) 3216827bd09bSSatish Balay { 3217827bd09bSSatish Balay gs_gop_vec_local_in_plus(gs,vals,step); 3218827bd09bSSatish Balay 3219827bd09bSSatish Balay /* pairwise */ 3220827bd09bSSatish Balay if (gs->num_pairs) 3221827bd09bSSatish Balay {gs_gop_vec_pairwise_plus(gs,vals,step);} 3222827bd09bSSatish Balay 3223827bd09bSSatish Balay /* tree */ 3224827bd09bSSatish Balay else if (gs->max_left_over) 3225827bd09bSSatish Balay {gs_gop_vec_tree_plus(gs,vals,step);} 3226827bd09bSSatish Balay 3227827bd09bSSatish Balay gs_gop_vec_local_out(gs,vals,step); 3228827bd09bSSatish Balay } 3229827bd09bSSatish Balay /* if intersection tree/pairwise and local is empty */ 3230827bd09bSSatish Balay else 3231827bd09bSSatish Balay { 3232827bd09bSSatish Balay /* pairwise */ 3233827bd09bSSatish Balay if (gs->num_pairs) 3234827bd09bSSatish Balay {gs_gop_vec_pairwise_plus(gs,vals,step);} 3235827bd09bSSatish Balay 3236827bd09bSSatish Balay /* tree */ 3237827bd09bSSatish Balay else if (gs->max_left_over) 3238827bd09bSSatish Balay {gs_gop_vec_tree_plus(gs,vals,step);} 3239827bd09bSSatish Balay } 32403fdc5746SBarry Smith PetscFunctionReturn(0); 3241827bd09bSSatish Balay } 3242827bd09bSSatish Balay 3243827bd09bSSatish Balay 3244827bd09bSSatish Balay 3245827bd09bSSatish Balay /****************************************************************************** 3246827bd09bSSatish Balay Function: gather_scatter 3247827bd09bSSatish Balay 3248827bd09bSSatish Balay Input : 3249827bd09bSSatish Balay Output: 3250827bd09bSSatish Balay Return: 3251827bd09bSSatish Balay Description: 3252827bd09bSSatish Balay ******************************************************************************/ 32530924e98cSBarry Smith static PetscErrorCode gs_gop_vec_local_plus( gs_id *gs, PetscScalar *vals, int step) 3254827bd09bSSatish Balay { 3255a501084fSBarry Smith int *num, *map, **reduce; 3256a501084fSBarry Smith PetscScalar *base; 3257827bd09bSSatish Balay 32583fdc5746SBarry Smith PetscFunctionBegin; 3259827bd09bSSatish Balay num = gs->num_local_reduce; 3260827bd09bSSatish Balay reduce = gs->local_reduce; 3261827bd09bSSatish Balay while ((map = *reduce)) 3262827bd09bSSatish Balay { 3263827bd09bSSatish Balay base = vals + map[0] * step; 3264827bd09bSSatish Balay 3265827bd09bSSatish Balay /* wall */ 3266827bd09bSSatish Balay if (*num == 2) 3267827bd09bSSatish Balay { 3268827bd09bSSatish Balay num++; reduce++; 3269827bd09bSSatish Balay rvec_add (base,vals+map[1]*step,step); 3270827bd09bSSatish Balay rvec_copy(vals+map[1]*step,base,step); 3271827bd09bSSatish Balay } 3272827bd09bSSatish Balay /* corner shared by three elements */ 3273827bd09bSSatish Balay else if (*num == 3) 3274827bd09bSSatish Balay { 3275827bd09bSSatish Balay num++; reduce++; 3276827bd09bSSatish Balay rvec_add (base,vals+map[1]*step,step); 3277827bd09bSSatish Balay rvec_add (base,vals+map[2]*step,step); 3278827bd09bSSatish Balay rvec_copy(vals+map[2]*step,base,step); 3279827bd09bSSatish Balay rvec_copy(vals+map[1]*step,base,step); 3280827bd09bSSatish Balay } 3281827bd09bSSatish Balay /* corner shared by four elements */ 3282827bd09bSSatish Balay else if (*num == 4) 3283827bd09bSSatish Balay { 3284827bd09bSSatish Balay num++; reduce++; 3285827bd09bSSatish Balay rvec_add (base,vals+map[1]*step,step); 3286827bd09bSSatish Balay rvec_add (base,vals+map[2]*step,step); 3287827bd09bSSatish Balay rvec_add (base,vals+map[3]*step,step); 3288827bd09bSSatish Balay rvec_copy(vals+map[3]*step,base,step); 3289827bd09bSSatish Balay rvec_copy(vals+map[2]*step,base,step); 3290827bd09bSSatish Balay rvec_copy(vals+map[1]*step,base,step); 3291827bd09bSSatish Balay } 3292827bd09bSSatish Balay /* general case ... odd geoms ... 3D */ 3293827bd09bSSatish Balay else 3294827bd09bSSatish Balay { 3295827bd09bSSatish Balay num++; 3296827bd09bSSatish Balay while (*++map >= 0) 3297827bd09bSSatish Balay {rvec_add (base,vals+*map*step,step);} 3298827bd09bSSatish Balay 3299827bd09bSSatish Balay map = *reduce; 3300827bd09bSSatish Balay while (*++map >= 0) 3301827bd09bSSatish Balay {rvec_copy(vals+*map*step,base,step);} 3302827bd09bSSatish Balay 3303827bd09bSSatish Balay reduce++; 3304827bd09bSSatish Balay } 3305827bd09bSSatish Balay } 33063fdc5746SBarry Smith PetscFunctionReturn(0); 3307827bd09bSSatish Balay } 3308827bd09bSSatish Balay 3309827bd09bSSatish Balay 3310827bd09bSSatish Balay 3311827bd09bSSatish Balay /****************************************************************************** 3312827bd09bSSatish Balay Function: gather_scatter 3313827bd09bSSatish Balay 3314827bd09bSSatish Balay Input : 3315827bd09bSSatish Balay Output: 3316827bd09bSSatish Balay Return: 3317827bd09bSSatish Balay Description: 3318827bd09bSSatish Balay ******************************************************************************/ 33190924e98cSBarry Smith static PetscErrorCode gs_gop_vec_local_in_plus( gs_id *gs, PetscScalar *vals, int step) 3320827bd09bSSatish Balay { 3321a501084fSBarry Smith int *num, *map, **reduce; 3322a501084fSBarry Smith PetscScalar *base; 33233fdc5746SBarry Smith PetscFunctionBegin; 3324827bd09bSSatish Balay num = gs->num_gop_local_reduce; 3325827bd09bSSatish Balay reduce = gs->gop_local_reduce; 3326827bd09bSSatish Balay while ((map = *reduce++)) 3327827bd09bSSatish Balay { 3328827bd09bSSatish Balay base = vals + map[0] * step; 3329827bd09bSSatish Balay 3330827bd09bSSatish Balay /* wall */ 3331827bd09bSSatish Balay if (*num == 2) 3332827bd09bSSatish Balay { 3333827bd09bSSatish Balay num ++; 3334827bd09bSSatish Balay rvec_add(base,vals+map[1]*step,step); 3335827bd09bSSatish Balay } 3336827bd09bSSatish Balay /* corner shared by three elements */ 3337827bd09bSSatish Balay else if (*num == 3) 3338827bd09bSSatish Balay { 3339827bd09bSSatish Balay num ++; 3340827bd09bSSatish Balay rvec_add(base,vals+map[1]*step,step); 3341827bd09bSSatish Balay rvec_add(base,vals+map[2]*step,step); 3342827bd09bSSatish Balay } 3343827bd09bSSatish Balay /* corner shared by four elements */ 3344827bd09bSSatish Balay else if (*num == 4) 3345827bd09bSSatish Balay { 3346827bd09bSSatish Balay num ++; 3347827bd09bSSatish Balay rvec_add(base,vals+map[1]*step,step); 3348827bd09bSSatish Balay rvec_add(base,vals+map[2]*step,step); 3349827bd09bSSatish Balay rvec_add(base,vals+map[3]*step,step); 3350827bd09bSSatish Balay } 3351827bd09bSSatish Balay /* general case ... odd geoms ... 3D*/ 3352827bd09bSSatish Balay else 3353827bd09bSSatish Balay { 3354827bd09bSSatish Balay num++; 3355827bd09bSSatish Balay while (*++map >= 0) 3356827bd09bSSatish Balay {rvec_add(base,vals+*map*step,step);} 3357827bd09bSSatish Balay } 3358827bd09bSSatish Balay } 33593fdc5746SBarry Smith PetscFunctionReturn(0); 3360827bd09bSSatish Balay } 3361827bd09bSSatish Balay 3362827bd09bSSatish Balay 3363827bd09bSSatish Balay /****************************************************************************** 3364827bd09bSSatish Balay Function: gather_scatter 3365827bd09bSSatish Balay 3366827bd09bSSatish Balay Input : 3367827bd09bSSatish Balay Output: 3368827bd09bSSatish Balay Return: 3369827bd09bSSatish Balay Description: 3370827bd09bSSatish Balay ******************************************************************************/ 33710924e98cSBarry Smith static PetscErrorCode gs_gop_vec_local_out( gs_id *gs, PetscScalar *vals, int step) 3372827bd09bSSatish Balay { 3373a501084fSBarry Smith int *num, *map, **reduce; 3374a501084fSBarry Smith PetscScalar *base; 3375827bd09bSSatish Balay 33763fdc5746SBarry Smith PetscFunctionBegin; 3377827bd09bSSatish Balay num = gs->num_gop_local_reduce; 3378827bd09bSSatish Balay reduce = gs->gop_local_reduce; 3379827bd09bSSatish Balay while ((map = *reduce++)) 3380827bd09bSSatish Balay { 3381827bd09bSSatish Balay base = vals + map[0] * step; 3382827bd09bSSatish Balay 3383827bd09bSSatish Balay /* wall */ 3384827bd09bSSatish Balay if (*num == 2) 3385827bd09bSSatish Balay { 3386827bd09bSSatish Balay num ++; 3387827bd09bSSatish Balay rvec_copy(vals+map[1]*step,base,step); 3388827bd09bSSatish Balay } 3389827bd09bSSatish Balay /* corner shared by three elements */ 3390827bd09bSSatish Balay else if (*num == 3) 3391827bd09bSSatish Balay { 3392827bd09bSSatish Balay num ++; 3393827bd09bSSatish Balay rvec_copy(vals+map[1]*step,base,step); 3394827bd09bSSatish Balay rvec_copy(vals+map[2]*step,base,step); 3395827bd09bSSatish Balay } 3396827bd09bSSatish Balay /* corner shared by four elements */ 3397827bd09bSSatish Balay else if (*num == 4) 3398827bd09bSSatish Balay { 3399827bd09bSSatish Balay num ++; 3400827bd09bSSatish Balay rvec_copy(vals+map[1]*step,base,step); 3401827bd09bSSatish Balay rvec_copy(vals+map[2]*step,base,step); 3402827bd09bSSatish Balay rvec_copy(vals+map[3]*step,base,step); 3403827bd09bSSatish Balay } 3404827bd09bSSatish Balay /* general case ... odd geoms ... 3D*/ 3405827bd09bSSatish Balay else 3406827bd09bSSatish Balay { 3407827bd09bSSatish Balay num++; 3408827bd09bSSatish Balay while (*++map >= 0) 3409827bd09bSSatish Balay {rvec_copy(vals+*map*step,base,step);} 3410827bd09bSSatish Balay } 3411827bd09bSSatish Balay } 34123fdc5746SBarry Smith PetscFunctionReturn(0); 3413827bd09bSSatish Balay } 3414827bd09bSSatish Balay 3415827bd09bSSatish Balay 3416827bd09bSSatish Balay 3417827bd09bSSatish Balay /****************************************************************************** 3418827bd09bSSatish Balay Function: gather_scatter 3419827bd09bSSatish Balay 3420827bd09bSSatish Balay VERSION 3 :: 3421827bd09bSSatish Balay 3422827bd09bSSatish Balay Input : 3423827bd09bSSatish Balay Output: 3424827bd09bSSatish Balay Return: 3425827bd09bSSatish Balay Description: 3426827bd09bSSatish Balay ******************************************************************************/ 34270924e98cSBarry Smith static PetscErrorCode gs_gop_vec_pairwise_plus( gs_id *gs, PetscScalar *in_vals, int step) 3428827bd09bSSatish Balay { 3429a501084fSBarry Smith PetscScalar *dptr1, *dptr2, *dptr3, *in1, *in2; 3430a501084fSBarry Smith int *iptr, *msg_list, *msg_size, **msg_nodes; 3431a501084fSBarry Smith int *pw, *list, *size, **nodes; 3432827bd09bSSatish Balay MPI_Request *msg_ids_in, *msg_ids_out, *ids_in, *ids_out; 3433827bd09bSSatish Balay MPI_Status status; 3434a501084fSBarry Smith PetscBLASInt i1; 34353fdc5746SBarry Smith PetscErrorCode ierr; 3436827bd09bSSatish Balay 34373fdc5746SBarry Smith PetscFunctionBegin; 3438a501084fSBarry Smith /* strip and load s */ 3439827bd09bSSatish Balay msg_list =list = gs->pair_list; 3440827bd09bSSatish Balay msg_size =size = gs->msg_sizes; 3441827bd09bSSatish Balay msg_nodes=nodes = gs->node_list; 3442827bd09bSSatish Balay iptr=pw = gs->pw_elm_list; 3443827bd09bSSatish Balay dptr1=dptr3 = gs->pw_vals; 3444827bd09bSSatish Balay msg_ids_in = ids_in = gs->msg_ids_in; 3445827bd09bSSatish Balay msg_ids_out = ids_out = gs->msg_ids_out; 3446827bd09bSSatish Balay dptr2 = gs->out; 3447827bd09bSSatish Balay in1=in2 = gs->in; 3448827bd09bSSatish Balay 3449827bd09bSSatish Balay /* post the receives */ 3450827bd09bSSatish Balay /* msg_nodes=nodes; */ 3451827bd09bSSatish Balay do 3452827bd09bSSatish Balay { 3453827bd09bSSatish Balay /* Should MPI_ANY_SOURCE be replaced by *list ? In that case do the 3454827bd09bSSatish Balay second one *list and do list++ afterwards */ 34553fdc5746SBarry Smith ierr = MPI_Irecv(in1, *size *step, MPIU_SCALAR, MPI_ANY_SOURCE, MSGTAG1 + *list++, gs->gs_comm, msg_ids_in++);CHKERRQ(ierr); 3456827bd09bSSatish Balay in1 += *size++ *step; 3457827bd09bSSatish Balay } 3458827bd09bSSatish Balay while (*++msg_nodes); 3459827bd09bSSatish Balay msg_nodes=nodes; 3460827bd09bSSatish Balay 3461827bd09bSSatish Balay /* load gs values into in out gs buffers */ 3462827bd09bSSatish Balay while (*iptr >= 0) 3463827bd09bSSatish Balay { 3464827bd09bSSatish Balay rvec_copy(dptr3,in_vals + *iptr*step,step); 3465827bd09bSSatish Balay dptr3+=step; 3466827bd09bSSatish Balay iptr++; 3467827bd09bSSatish Balay } 3468827bd09bSSatish Balay 3469827bd09bSSatish Balay /* load out buffers and post the sends */ 3470827bd09bSSatish Balay while ((iptr = *msg_nodes++)) 3471827bd09bSSatish Balay { 3472827bd09bSSatish Balay dptr3 = dptr2; 3473827bd09bSSatish Balay while (*iptr >= 0) 3474827bd09bSSatish Balay { 3475827bd09bSSatish Balay rvec_copy(dptr2,dptr1 + *iptr*step,step); 3476827bd09bSSatish Balay dptr2+=step; 3477827bd09bSSatish Balay iptr++; 3478827bd09bSSatish Balay } 34793fdc5746SBarry Smith ierr = MPI_Isend(dptr3, *msg_size++ *step, MPIU_SCALAR, *msg_list++, MSGTAG1+my_id, gs->gs_comm, msg_ids_out++);CHKERRQ(ierr); 3480827bd09bSSatish Balay } 3481827bd09bSSatish Balay 3482827bd09bSSatish Balay /* tree */ 3483827bd09bSSatish Balay if (gs->max_left_over) 3484827bd09bSSatish Balay {gs_gop_vec_tree_plus(gs,in_vals,step);} 3485827bd09bSSatish Balay 3486827bd09bSSatish Balay /* process the received data */ 3487827bd09bSSatish Balay msg_nodes=nodes; 3488a501084fSBarry Smith while ((iptr = *nodes++)){ 3489a501084fSBarry Smith PetscScalar d1 = 1.0; 3490827bd09bSSatish Balay /* Should I check the return value of MPI_Wait() or status? */ 3491827bd09bSSatish Balay /* Can this loop be replaced by a call to MPI_Waitall()? */ 34923fdc5746SBarry Smith ierr = MPI_Wait(ids_in++, &status);CHKERRQ(ierr); 3493a501084fSBarry Smith while (*iptr >= 0) { 349471044d3cSBarry Smith BLASaxpy_(&step,&d1,in2,&i1,dptr1 + *iptr*step,&i1); 3495827bd09bSSatish Balay in2+=step; 3496827bd09bSSatish Balay iptr++; 3497827bd09bSSatish Balay } 3498827bd09bSSatish Balay } 3499827bd09bSSatish Balay 3500827bd09bSSatish Balay /* replace vals */ 3501827bd09bSSatish Balay while (*pw >= 0) 3502827bd09bSSatish Balay { 3503827bd09bSSatish Balay rvec_copy(in_vals + *pw*step,dptr1,step); 3504827bd09bSSatish Balay dptr1+=step; 3505827bd09bSSatish Balay pw++; 3506827bd09bSSatish Balay } 3507827bd09bSSatish Balay 3508827bd09bSSatish Balay /* clear isend message handles */ 3509827bd09bSSatish Balay /* This changed for clarity though it could be the same */ 3510827bd09bSSatish Balay while (*msg_nodes++) 3511827bd09bSSatish Balay /* Should I check the return value of MPI_Wait() or status? */ 3512827bd09bSSatish Balay /* Can this loop be replaced by a call to MPI_Waitall()? */ 35133fdc5746SBarry Smith {ierr = MPI_Wait(ids_out++, &status);CHKERRQ(ierr);} 3514827bd09bSSatish Balay 35153fdc5746SBarry Smith PetscFunctionReturn(0); 3516827bd09bSSatish Balay } 3517827bd09bSSatish Balay 3518827bd09bSSatish Balay 3519827bd09bSSatish Balay 3520827bd09bSSatish Balay /****************************************************************************** 3521827bd09bSSatish Balay Function: gather_scatter 3522827bd09bSSatish Balay 3523827bd09bSSatish Balay Input : 3524827bd09bSSatish Balay Output: 3525827bd09bSSatish Balay Return: 3526827bd09bSSatish Balay Description: 3527827bd09bSSatish Balay ******************************************************************************/ 35280924e98cSBarry Smith static PetscErrorCode gs_gop_vec_tree_plus( gs_id *gs, PetscScalar *vals, int step) 3529827bd09bSSatish Balay { 3530a501084fSBarry Smith int size, *in, *out; 3531a501084fSBarry Smith PetscScalar *buf, *work; 3532827bd09bSSatish Balay int op[] = {GL_ADD,0}; 3533a501084fSBarry Smith PetscBLASInt i1 = 1; 3534827bd09bSSatish Balay 35353fdc5746SBarry Smith PetscFunctionBegin; 3536827bd09bSSatish Balay /* copy over to local variables */ 3537827bd09bSSatish Balay in = gs->tree_map_in; 3538827bd09bSSatish Balay out = gs->tree_map_out; 3539827bd09bSSatish Balay buf = gs->tree_buf; 3540827bd09bSSatish Balay work = gs->tree_work; 3541827bd09bSSatish Balay size = gs->tree_nel*step; 3542827bd09bSSatish Balay 3543827bd09bSSatish Balay /* zero out collection buffer */ 3544827bd09bSSatish Balay rvec_zero(buf,size); 3545827bd09bSSatish Balay 3546827bd09bSSatish Balay 3547827bd09bSSatish Balay /* copy over my contributions */ 3548827bd09bSSatish Balay while (*in >= 0) 3549827bd09bSSatish Balay { 355071044d3cSBarry Smith BLAScopy_(&step,vals + *in++*step,&i1,buf + *out++*step,&i1); 3551827bd09bSSatish Balay } 3552827bd09bSSatish Balay 3553827bd09bSSatish Balay /* perform fan in/out on full buffer */ 3554827bd09bSSatish Balay /* must change grop to handle the blas */ 3555827bd09bSSatish Balay grop(buf,work,size,op); 3556827bd09bSSatish Balay 3557827bd09bSSatish Balay /* reset */ 3558827bd09bSSatish Balay in = gs->tree_map_in; 3559827bd09bSSatish Balay out = gs->tree_map_out; 3560827bd09bSSatish Balay 3561827bd09bSSatish Balay /* get the portion of the results I need */ 3562827bd09bSSatish Balay while (*in >= 0) 3563827bd09bSSatish Balay { 356471044d3cSBarry Smith BLAScopy_(&step,buf + *out++*step,&i1,vals + *in++*step,&i1); 3565827bd09bSSatish Balay } 35663fdc5746SBarry Smith PetscFunctionReturn(0); 3567827bd09bSSatish Balay } 3568827bd09bSSatish Balay 3569827bd09bSSatish Balay 3570827bd09bSSatish Balay 3571827bd09bSSatish Balay /****************************************************************************** 3572827bd09bSSatish Balay Function: gather_scatter 3573827bd09bSSatish Balay 3574827bd09bSSatish Balay Input : 3575827bd09bSSatish Balay Output: 3576827bd09bSSatish Balay Return: 3577827bd09bSSatish Balay Description: 3578827bd09bSSatish Balay ******************************************************************************/ 35790924e98cSBarry Smith PetscErrorCode gs_gop_hc( gs_id *gs, PetscScalar *vals, const char *op, int dim) 3580827bd09bSSatish Balay { 35813fdc5746SBarry Smith PetscFunctionBegin; 3582827bd09bSSatish Balay switch (*op) { 3583827bd09bSSatish Balay case '+': 3584827bd09bSSatish Balay gs_gop_plus_hc(gs,vals,dim); 3585827bd09bSSatish Balay break; 3586827bd09bSSatish Balay default: 3587827bd09bSSatish Balay error_msg_warning("gs_gop_hc() :: %c is not a valid op",op[0]); 3588827bd09bSSatish Balay error_msg_warning("gs_gop_hc() :: default :: plus\n"); 3589827bd09bSSatish Balay gs_gop_plus_hc(gs,vals,dim); 3590827bd09bSSatish Balay break; 3591827bd09bSSatish Balay } 35923fdc5746SBarry Smith PetscFunctionReturn(0); 3593827bd09bSSatish Balay } 3594827bd09bSSatish Balay 3595827bd09bSSatish Balay 3596827bd09bSSatish Balay 3597827bd09bSSatish Balay /****************************************************************************** 3598827bd09bSSatish Balay Function: gather_scatter 3599827bd09bSSatish Balay 3600827bd09bSSatish Balay Input : 3601827bd09bSSatish Balay Output: 3602827bd09bSSatish Balay Return: 3603827bd09bSSatish Balay Description: 3604827bd09bSSatish Balay ******************************************************************************/ 36050924e98cSBarry Smith static PetscErrorCode gs_gop_plus_hc( gs_id *gs, PetscScalar *vals, int dim) 3606827bd09bSSatish Balay { 36073fdc5746SBarry Smith PetscFunctionBegin; 3608827bd09bSSatish Balay /* if there's nothing to do return */ 3609827bd09bSSatish Balay if (dim<=0) 36103fdc5746SBarry Smith { PetscFunctionReturn(0);} 3611827bd09bSSatish Balay 3612827bd09bSSatish Balay /* can't do more dimensions then exist */ 361339945688SSatish Balay dim = PetscMin(dim,i_log2_num_nodes); 3614827bd09bSSatish Balay 3615827bd09bSSatish Balay /* local only operations!!! */ 3616827bd09bSSatish Balay if (gs->num_local) 3617827bd09bSSatish Balay {gs_gop_local_plus(gs,vals);} 3618827bd09bSSatish Balay 3619827bd09bSSatish Balay /* if intersection tree/pairwise and local isn't empty */ 3620827bd09bSSatish Balay if (gs->num_local_gop) 3621827bd09bSSatish Balay { 3622827bd09bSSatish Balay gs_gop_local_in_plus(gs,vals); 3623827bd09bSSatish Balay 3624827bd09bSSatish Balay /* pairwise will do tree inside ... */ 3625827bd09bSSatish Balay if (gs->num_pairs) 3626827bd09bSSatish Balay {gs_gop_pairwise_plus_hc(gs,vals,dim);} 3627827bd09bSSatish Balay 3628827bd09bSSatish Balay /* tree only */ 3629827bd09bSSatish Balay else if (gs->max_left_over) 3630827bd09bSSatish Balay {gs_gop_tree_plus_hc(gs,vals,dim);} 3631827bd09bSSatish Balay 3632827bd09bSSatish Balay gs_gop_local_out(gs,vals); 3633827bd09bSSatish Balay } 3634827bd09bSSatish Balay /* if intersection tree/pairwise and local is empty */ 3635827bd09bSSatish Balay else 3636827bd09bSSatish Balay { 3637827bd09bSSatish Balay /* pairwise will do tree inside */ 3638827bd09bSSatish Balay if (gs->num_pairs) 3639827bd09bSSatish Balay {gs_gop_pairwise_plus_hc(gs,vals,dim);} 3640827bd09bSSatish Balay 3641827bd09bSSatish Balay /* tree */ 3642827bd09bSSatish Balay else if (gs->max_left_over) 3643827bd09bSSatish Balay {gs_gop_tree_plus_hc(gs,vals,dim);} 3644827bd09bSSatish Balay } 36453fdc5746SBarry Smith PetscFunctionReturn(0); 3646827bd09bSSatish Balay } 3647827bd09bSSatish Balay 3648827bd09bSSatish Balay 3649827bd09bSSatish Balay /****************************************************************************** 3650827bd09bSSatish Balay VERSION 3 :: 3651827bd09bSSatish Balay 3652827bd09bSSatish Balay Input : 3653827bd09bSSatish Balay Output: 3654827bd09bSSatish Balay Return: 3655827bd09bSSatish Balay Description: 3656827bd09bSSatish Balay ******************************************************************************/ 36570924e98cSBarry Smith static PetscErrorCode gs_gop_pairwise_plus_hc( gs_id *gs, PetscScalar *in_vals, int dim) 3658827bd09bSSatish Balay { 3659a501084fSBarry Smith PetscScalar *dptr1, *dptr2, *dptr3, *in1, *in2; 3660a501084fSBarry Smith int *iptr, *msg_list, *msg_size, **msg_nodes; 3661a501084fSBarry Smith int *pw, *list, *size, **nodes; 3662827bd09bSSatish Balay MPI_Request *msg_ids_in, *msg_ids_out, *ids_in, *ids_out; 3663827bd09bSSatish Balay MPI_Status status; 3664827bd09bSSatish Balay int i, mask=1; 36653fdc5746SBarry Smith PetscErrorCode ierr; 3666827bd09bSSatish Balay 36673fdc5746SBarry Smith PetscFunctionBegin; 3668827bd09bSSatish Balay for (i=1; i<dim; i++) 3669827bd09bSSatish Balay {mask<<=1; mask++;} 3670827bd09bSSatish Balay 3671827bd09bSSatish Balay 3672a501084fSBarry Smith /* strip and load s */ 3673827bd09bSSatish Balay msg_list =list = gs->pair_list; 3674827bd09bSSatish Balay msg_size =size = gs->msg_sizes; 3675827bd09bSSatish Balay msg_nodes=nodes = gs->node_list; 3676827bd09bSSatish Balay iptr=pw = gs->pw_elm_list; 3677827bd09bSSatish Balay dptr1=dptr3 = gs->pw_vals; 3678827bd09bSSatish Balay msg_ids_in = ids_in = gs->msg_ids_in; 3679827bd09bSSatish Balay msg_ids_out = ids_out = gs->msg_ids_out; 3680827bd09bSSatish Balay dptr2 = gs->out; 3681827bd09bSSatish Balay in1=in2 = gs->in; 3682827bd09bSSatish Balay 3683827bd09bSSatish Balay /* post the receives */ 3684827bd09bSSatish Balay /* msg_nodes=nodes; */ 3685827bd09bSSatish Balay do 3686827bd09bSSatish Balay { 3687827bd09bSSatish Balay /* Should MPI_ANY_SOURCE be replaced by *list ? In that case do the 3688827bd09bSSatish Balay second one *list and do list++ afterwards */ 3689827bd09bSSatish Balay if ((my_id|mask)==(*list|mask)) 3690827bd09bSSatish Balay { 36913fdc5746SBarry Smith ierr = MPI_Irecv(in1, *size, MPIU_SCALAR, MPI_ANY_SOURCE, MSGTAG1 + *list++, gs->gs_comm, msg_ids_in++);CHKERRQ(ierr); 3692827bd09bSSatish Balay in1 += *size++; 3693827bd09bSSatish Balay } 3694827bd09bSSatish Balay else 3695827bd09bSSatish Balay {list++; size++;} 3696827bd09bSSatish Balay } 3697827bd09bSSatish Balay while (*++msg_nodes); 3698827bd09bSSatish Balay 3699827bd09bSSatish Balay /* load gs values into in out gs buffers */ 3700827bd09bSSatish Balay while (*iptr >= 0) 3701827bd09bSSatish Balay {*dptr3++ = *(in_vals + *iptr++);} 3702827bd09bSSatish Balay 3703827bd09bSSatish Balay /* load out buffers and post the sends */ 3704827bd09bSSatish Balay msg_nodes=nodes; 3705827bd09bSSatish Balay list = msg_list; 3706827bd09bSSatish Balay while ((iptr = *msg_nodes++)) 3707827bd09bSSatish Balay { 3708827bd09bSSatish Balay if ((my_id|mask)==(*list|mask)) 3709827bd09bSSatish Balay { 3710827bd09bSSatish Balay dptr3 = dptr2; 3711827bd09bSSatish Balay while (*iptr >= 0) 3712827bd09bSSatish Balay {*dptr2++ = *(dptr1 + *iptr++);} 3713827bd09bSSatish Balay /* CHECK PERSISTENT COMMS MODE FOR ALL THIS STUFF */ 3714827bd09bSSatish Balay /* is msg_ids_out++ correct? */ 37153fdc5746SBarry Smith ierr = MPI_Isend(dptr3, *msg_size++, MPIU_SCALAR, *list++, MSGTAG1+my_id, gs->gs_comm, msg_ids_out++);CHKERRQ(ierr); 3716827bd09bSSatish Balay } 3717827bd09bSSatish Balay else 3718827bd09bSSatish Balay {list++; msg_size++;} 3719827bd09bSSatish Balay } 3720827bd09bSSatish Balay 3721827bd09bSSatish Balay /* do the tree while we're waiting */ 3722827bd09bSSatish Balay if (gs->max_left_over) 3723827bd09bSSatish Balay {gs_gop_tree_plus_hc(gs,in_vals,dim);} 3724827bd09bSSatish Balay 3725827bd09bSSatish Balay /* process the received data */ 3726827bd09bSSatish Balay msg_nodes=nodes; 3727827bd09bSSatish Balay list = msg_list; 3728827bd09bSSatish Balay while ((iptr = *nodes++)) 3729827bd09bSSatish Balay { 3730827bd09bSSatish Balay if ((my_id|mask)==(*list|mask)) 3731827bd09bSSatish Balay { 3732827bd09bSSatish Balay /* Should I check the return value of MPI_Wait() or status? */ 3733827bd09bSSatish Balay /* Can this loop be replaced by a call to MPI_Waitall()? */ 37343fdc5746SBarry Smith ierr = MPI_Wait(ids_in++, &status);CHKERRQ(ierr); 3735827bd09bSSatish Balay while (*iptr >= 0) 3736827bd09bSSatish Balay {*(dptr1 + *iptr++) += *in2++;} 3737827bd09bSSatish Balay } 3738827bd09bSSatish Balay list++; 3739827bd09bSSatish Balay } 3740827bd09bSSatish Balay 3741827bd09bSSatish Balay /* replace vals */ 3742827bd09bSSatish Balay while (*pw >= 0) 3743827bd09bSSatish Balay {*(in_vals + *pw++) = *dptr1++;} 3744827bd09bSSatish Balay 3745827bd09bSSatish Balay /* clear isend message handles */ 3746827bd09bSSatish Balay /* This changed for clarity though it could be the same */ 3747827bd09bSSatish Balay while (*msg_nodes++) 3748827bd09bSSatish Balay { 3749827bd09bSSatish Balay if ((my_id|mask)==(*msg_list|mask)) 3750827bd09bSSatish Balay { 3751827bd09bSSatish Balay /* Should I check the return value of MPI_Wait() or status? */ 3752827bd09bSSatish Balay /* Can this loop be replaced by a call to MPI_Waitall()? */ 37533fdc5746SBarry Smith ierr = MPI_Wait(ids_out++, &status);CHKERRQ(ierr); 3754827bd09bSSatish Balay } 3755827bd09bSSatish Balay msg_list++; 3756827bd09bSSatish Balay } 3757827bd09bSSatish Balay 37583fdc5746SBarry Smith PetscFunctionReturn(0); 3759827bd09bSSatish Balay } 3760827bd09bSSatish Balay 3761827bd09bSSatish Balay 3762827bd09bSSatish Balay 3763827bd09bSSatish Balay /****************************************************************************** 3764827bd09bSSatish Balay Function: gather_scatter 3765827bd09bSSatish Balay 3766827bd09bSSatish Balay Input : 3767827bd09bSSatish Balay Output: 3768827bd09bSSatish Balay Return: 3769827bd09bSSatish Balay Description: 3770827bd09bSSatish Balay ******************************************************************************/ 37710924e98cSBarry Smith static PetscErrorCode gs_gop_tree_plus_hc(gs_id *gs, PetscScalar *vals, int dim) 3772827bd09bSSatish Balay { 3773827bd09bSSatish Balay int size; 3774827bd09bSSatish Balay int *in, *out; 3775a501084fSBarry Smith PetscScalar *buf, *work; 3776827bd09bSSatish Balay int op[] = {GL_ADD,0}; 3777827bd09bSSatish Balay 37783fdc5746SBarry Smith PetscFunctionBegin; 3779827bd09bSSatish Balay in = gs->tree_map_in; 3780827bd09bSSatish Balay out = gs->tree_map_out; 3781827bd09bSSatish Balay buf = gs->tree_buf; 3782827bd09bSSatish Balay work = gs->tree_work; 3783827bd09bSSatish Balay size = gs->tree_nel; 3784827bd09bSSatish Balay 3785827bd09bSSatish Balay rvec_zero(buf,size); 3786827bd09bSSatish Balay 3787827bd09bSSatish Balay while (*in >= 0) 3788827bd09bSSatish Balay {*(buf + *out++) = *(vals + *in++);} 3789827bd09bSSatish Balay 3790827bd09bSSatish Balay in = gs->tree_map_in; 3791827bd09bSSatish Balay out = gs->tree_map_out; 3792827bd09bSSatish Balay 3793827bd09bSSatish Balay grop_hc(buf,work,size,op,dim); 3794827bd09bSSatish Balay 3795827bd09bSSatish Balay while (*in >= 0) 3796827bd09bSSatish Balay {*(vals + *in++) = *(buf + *out++);} 37973fdc5746SBarry Smith PetscFunctionReturn(0); 3798827bd09bSSatish Balay } 3799827bd09bSSatish Balay 3800827bd09bSSatish Balay 3801827bd09bSSatish Balay 3802