wrfv2_fire/external/RSL/RSL/rsl.h

   1 /***********************************************************************
   2
   3                               COPYRIGHT
   4
   5      The following is a notice of limited availability of the code and
   6      Government license and disclaimer which must be included in the
   7      prologue of the code and in all source listings of the code.
   8
   9      Copyright notice
  10        (c) 1977  University of Chicago
  11
  12      Permission is hereby granted to use, reproduce, prepare
  13      derivative works, and to redistribute to others at no charge.  If
  14      you distribute a copy or copies of the Software, or you modify a
  15      copy or copies of the Software or any portion of it, thus forming
  16      a work based on the Software and make and/or distribute copies of
  17      such work, you must meet the following conditions:
  18
  19           a) If you make a copy of the Software (modified or verbatim)
  20              it must include the copyright notice and Government
  21              license and disclaimer.
  22
  23           b) You must cause the modified Software to carry prominent
  24              notices stating that you changed specified portions of
  25              the Software.
  26
  27      This software was authored by:
  28
  29      Argonne National Laboratory
  30      J. Michalakes: (630) 252-6646; email: michalak@mcs.anl.gov
  31      Mathematics and Computer Science Division
  32      Argonne National Laboratory, Argonne, IL  60439
  33
  34      ARGONNE NATIONAL LABORATORY (ANL), WITH FACILITIES IN THE STATES
  35      OF ILLINOIS AND IDAHO, IS OWNED BY THE UNITED STATES GOVERNMENT,
  36      AND OPERATED BY THE UNIVERSITY OF CHICAGO UNDER PROVISION OF A
  37      CONTRACT WITH THE DEPARTMENT OF ENERGY.
  38
  39                       GOVERNMENT LICENSE AND DISCLAIMER
  40
  41      This computer code material was prepared, in part, as an account
  42      of work sponsored by an agency of the United States Government.
  43      The Government is granted for itself and others acting on its
  44      behalf a paid-up, nonexclusive, irrevocable worldwide license in
  45      this data to reproduce, prepare derivative works, distribute
  46      copies to the public, perform publicly and display publicly, and
  47      to permit others to do so.  NEITHER THE UNITED STATES GOVERNMENT
  48      NOR ANY AGENCY THEREOF, NOR THE UNIVERSITY OF CHICAGO, NOR ANY OF
  49      THEIR EMPLOYEES, MAKES ANY WARRANTY, EXPRESS OR IMPLIED, OR
  50      ASSUMES ANY LEGAL LIABILITY OR RESPONSIBILITY FOR THE ACCURACY,
  51      COMPLETENESS, OR USEFULNESS OF ANY INFORMATION, APPARATUS,
  52      PRODUCT, OR PROCESS DISCLOSED, OR REPRESENTS THAT ITS USE WOULD
  53      NOT INFRINGE PRIVATELY OWNED RIGHTS.
  54
  55 ***************************************************************************/
  56
  57
  58 #ifndef RSL_H
  59 #define RSL_H
  60
  61 /**************************************************************/
  62 /* Any changes to this file should be followed by make clean  */
  63 /* before remaking the RSL library                            */
  64 /**************************************************************/
  65
  66 /**************************************************************/
  67 /* not likely that you will need to change things below here */
  68
  69 /* this definition used in cd.c */
  70 #define HARD_CODED_BOUNDARY_WIDTH_FIX_ME_PLEASE 5
  71
  72 #define RSL_MAXDOMAINS  MAXDOM_MAKE     /* FORTRAN */
  73
  74 #define RSL_MAXPROC     MAXPROC_MAKE    /* FORTRAN */
  75
  76 /* increased from 5 to 7 10/4/96 to accomodate 168 pt stencil */
  77 /* decreased to 3 (2 + 1 for garbage row) to save memory -- only estore
  78    in sound needs 168 pt stencil and we'll handle that as a special case */
  79 #define RSL_DEFAULT_PADAREA 4           /* FORTRAN */
  80
  81 #define RSL_MLOW        1               /* FORTRAN */
  82 #define RSL_MHIGH       2               /* FORTRAN */
  83 #define RSL_NLOW        3               /* FORTRAN */
  84 #define RSL_NHIGH       4               /* FORTRAN */
  85 #define RSL_DBDY        5               /* FORTRAN */
  86 #define RSL_CLOSEST     6               /* FORTRAN */
  87
  88 #define DOT_BDY_INFO_LEN 10             /* FORTRAN */
  89
  90 #define RSL_MLOW_X      7               /* FORTRAN */
  91 #define RSL_MHIGH_X     8               /* FORTRAN */
  92 #define RSL_NLOW_X      9               /* FORTRAN */
  93 #define RSL_NHIGH_X    10               /* FORTRAN */
  94 #define RSL_DBDY_X     11               /* FORTRAN */
  95 #define RSL_CLOSEST_X  12               /* FORTRAN */
  96
  97 #define CROSS_BDY_INFO_LEN 12           /* FORTRAN */
  98
  99 #define RSL_00          7               /* FORTRAN */
 100 #define RSL_M0          8               /* FORTRAN */
 101 #define RSL_0N          9               /* FORTRAN */
 102 #define RSL_MN         10               /* FORTRAN */
 103
 104 #define MLOW (RSL_MLOW-1)
 105 #define MHIGH (RSL_MHIGH-1)
 106 #define NLOW (RSL_NLOW-1)
 107 #define NHIGH (RSL_NHIGH-1)
 108
 109 #define MAX_RUNPAD  6
 110 #define MAX_KINDPAD 8
 111 /* break ties in boundary classification */
 112 #define M_WINS      0                   /* FORTRAN */
 113 #define N_WINS      1                   /* FORTRAN */
 114 #define DIAG_WINS  2                   /* FORTRAN */
 115
 116 #ifdef SHOW_RSL_CONFIGURATION
 117 /* SHOW_RSL_CONFIGURATION is only every defined externally.  This
 118    next bit is never actually used.  Rather it is used by the top
 119    level makefile to show the configuration of the model when the
 120    user types make showconfig.  */
 121 show config:  ---if the library were completely remade---
 122 show config:"RSL_IMAX" would be set to:            RSL_IMAX
 123 show config:"RSL_JMAX" would be set to:            RSL_JMAX
 124 show config:"RSL_MAXDOMAINS" would be set to:      RSL_MAXDOMAINS
 125 show config:"RSL_MAXPROC" would be set to:         RSL_MAXPROC
 126 show config:  ---See also the makefile---
 127 #endif
 128
 129
 130 /* this should become runtime setaable from the fortran someday */
 131
 132 #define RSL_TRUE  1                     /* FORTRAN */
 133 #define RSL_FALSE 0                     /* FORTRAN */
 134 #define RSL_INVALID -1                  /* FORTRAN */
 135 #define RSL_VALID 1                     /* FORTRAN */
 136
 137 #define RSL_IRAX 3                      /* FORTRAN */
 138 #define RSL_MAXCHILDREN (RSL_IRAX*RSL_IRAX)
 139 #define RSL_MAXKIDS RSL_MAXCHILDREN
 140 #define RSL_MAXDESCRIPTORS  2048
 141 /* settings below will allow for 20000 fields */
 142 #define MAX_BASE_TABLE_ENTRIES 20000
 143
 144 #if (ALLOW_RSL_168PT == 1)
 145 #  define RSL_MAXSTEN    168              /* MAX NUMBER OF STENCIL PTS */
 146 #else
 147 #  define RSL_MAXSTEN    24             /* MAX NUMBER OF STENCIL PTS */
 148 #endif
 149
 150
 151 #define RSL_REAL                0       /* FORTRAN */
 152 #define RSL_DOUBLE              1       /* FORTRAN */
 153 #define RSL_COMPLEX             2       /* FORTRAN */
 154 #define RSL_INTEGER             3       /* FORTRAN */
 155 #define RSL_CHARACTER           4       /* FORTRAN */
 156
 157 #define RSL_REAL_F90            100     /* FORTRAN */
 158 #define RSL_DOUBLE_F90          101     /* FORTRAN */
 159 #define RSL_COMPLEX_F90         102     /* FORTRAN */
 160 #define RSL_INTEGER_F90         103     /* FORTRAN */
 161 #define RSL_CHARACTER_F90       104     /* FORTRAN */
 162
 163 /* traversal orders */
 164 /* A = Ascending, D = Descending, applied to dim in order left to right */
 165 #define MINMAJ_AA                0       /* FORTRAN */
 166 #define MINMAJ_AD                1       /* FORTRAN */
 167 #define MINMAJ_DA                2       /* FORTRAN */
 168 #define MINMAJ_DD                3       /* FORTRAN */
 169 #define MAJMIN_AA             1000       /* FORTRAN */
 170 #define MAJMIN_AD             1001       /* FORTRAN */
 171 #define MAJMIN_DA             1002       /* FORTRAN */
 172 #define MAJMIN_DD             1003       /* FORTRAN */
 173 #define MINMAJ           MINMAJ_AA       /* FORTRAN */
 174 #define MAJMIN           MAJMIN_AA       /* FORTRAN */
 175
 176 #define MNMJ(X)     ( ( (X) / 1000 ) == 0 )
 177 #define D1(X)       ( 2 & (X) )
 178 #define D2(X)       ( 1 & (X) )
 179 #define A1(X)       ( ! A1 )
 180 #define A2(X)       ( ! A2 )
 181
 182 /* nested domain shapes */
 183 #define RSL_REGULAR_NEST  55            /* FORTRAN */
 184 #define RSL_RAGGED_NEST  56             /* FORTRAN */
 185
 186 /* packing strategies */
 187 #define MINNS_MAJEW_2D   0
 188 #define MINEW_MAJNS_2D   1
 189 #define MINNS_MAJEW_K_3D 2
 190 #define MINEW_MAJNS_K_3D 3
 191 #define K_MIDNS_MAJEW_3D 4
 192 #define MINNS_K_MAJEW_3D 5
 193
 194 /* io strategies */
 195
 196 #define IO_REPL  99             /* FORTRAN */
 197
 198 #define IO2D_IJ  0              /* FORTRAN */
 199 #define IO2D_JI  1              /* FORTRAN */
 200 #define IO3D_IJK 2              /* FORTRAN */
 201 #define IO3D_JIK 3              /* FORTRAN */
 202 #define IO2D     4              /* FORTRAN */
 203 #define IO3D     5              /* FORTRAN */
 204 #define IO3D_KIJ 6              /* FORTRAN */
 205 #define IO3D_IKJ 7              /* FORTRAN */
 206
 207 #define IO2D_IJ_RAW  10         /* FORTRAN */
 208 #define IO2D_JI_RAW  11         /* FORTRAN */
 209 #define IO3D_IJK_RAW 12         /* FORTRAN */
 210 #define IO3D_JIK_RAW 13         /* FORTRAN */
 211
 212 #define IO2D_IJ_PORTAL  20      /* FORTRAN */
 213 #define IO2D_JI_PORTAL  21      /* FORTRAN */
 214 #define IO3D_IJK_PORTAL 22      /* FORTRAN */
 215 #define IO3D_JIK_PORTAL 23      /* FORTRAN */
 216
 217 #define IO2D_IJ_INTERNAL  24    /* FORTRAN */
 218 #define IO2D_JI_INTERNAL  25    /* FORTRAN */
 219 #define IO3D_IJK_INTERNAL 26    /* FORTRAN */
 220 #define IO3D_JIK_INTERNAL 27    /* FORTRAN */
 221 #define IO3D_KIJ_INTERNAL 28    /* FORTRAN */
 222 #define IO3D_IKJ_INTERNAL 29    /* FORTRAN */
 223
 224 #define IO2D_IJ_88  30  /* FORTRAN */
 225 #define IO2D_JI_88  31  /* FORTRAN */
 226 #define IO3D_IJK_88 32  /* FORTRAN */
 227 #define IO3D_JIK_88 33  /* FORTRAN */
 228
 229 #define RSL_MAXDIM 3
 230
 231 /* type declarations */
 232
 233 typedef int * int_p ;
 234 #if !(defined(SUNDEBUG) || defined(crayx1))
 235 #ifndef NEC_TUNE
 236 typedef short   rsl_processor_t ;
 237 typedef short   rsl_index_t ;
 238 typedef short   rsl_dimlen_t ;
 239 #else
 240 typedef int   rsl_processor_t ;
 241 typedef int   rsl_index_t ;
 242 typedef int   rsl_dimlen_t ;
 243 #endif
 244 #else
 245 typedef int   rsl_processor_t ;
 246 typedef int   rsl_index_t ;
 247 typedef int   rsl_dimlen_t ;
 248 #endif
 249 typedef long    rsl_point_id_t ;
 250
 251 #ifdef crayx1
 252 typedef int rsl_tag_t ;
 253 #else
 254 #ifndef NEC_TUNE
 255 typedef unsigned char rsl_tag_t ;
 256 #else
 257 typedef int rsl_tag_t ;
 258 #endif
 259 #endif
 260
 261 typedef struct rsl_list {
 262   struct rsl_list * next ;
 263   void * data ;                 /* pointer to some node */
 264 #ifdef crayx1
 265   int info1 ;                   /* blank info field */
 266   int info2 ;                   /* blank info field */
 267 #else
 268   short info1 ;                 /* blank info field */
 269   short info2 ;                 /* blank info field */
 270 #endif
 271 } rsl_list_t ;
 272
 273
 274 typedef struct rsl_runrec {     /* added 1/9/95 for rsl_compute_islab */
 275   int i, j, ig, jg, runlength ;
 276 } rsl_runrec_t ;
 277
 278 typedef struct bcast_point_desc {
 279   rsl_point_id_t nest_id ;      /* pt in nest */
 280   rsl_point_id_t parent_id ;    /* cd pt */
 281   unsigned char cn, cm ;        /* indices of point, with respect to cd pt */
 282 } bcast_point_desc_t ;
 283
 284 typedef bcast_point_desc_t merge_point_desc_t ;
 285
 286 typedef struct rsl_fldspec {
 287   struct rsl_fldspec    * next ;
 288   void                  * base ;
 289   rsl_tag_t             ndim ;
 290   rsl_tag_t             elemsz ;
 291   rsl_tag_t             memsize ;
 292   rsl_tag_t             type ;
 293   int                   f90_table_index ;
 294   rsl_tag_t             strategy ;
 295   rsl_tag_t             decomp[ RSL_MAXDIM ] ;
 296   rsl_index_t           gdex[ RSL_MAXDIM ] ;
 297   rsl_dimlen_t          glen[ RSL_MAXDIM ] ;
 298   rsl_dimlen_t          llen[ RSL_MAXDIM ] ;
 299   rsl_dimlen_t          stag[ RSL_MAXDIM ] ;  /* 0 = not staggered; 1 = staggered */
 300 } rsl_fldspec_t ;
 301
 302 typedef struct message_desc {
 303   rsl_tag_t tag;
 304     /* should be MESSAGE_DESC or BLANK_MESSAGE_DESC */
 305   int mh ;              /* handle */
 306   int nflds ;           /* number of field specs in message */
 307   int nbytes ;          /* for blank messages */
 308   rsl_fldspec_t *fldspecs ;
 309 } message_desc_t ;
 310
 311 typedef struct rsl_plist_elem {
 312   struct rsl_plist_elem *next ;
 313   rsl_point_id_t id ;
 314   rsl_processor_t P ;
 315 } rsl_plist_elem_t ;
 316
 317 typedef struct rsl_child_info {
 318   rsl_point_id_t  child[RSL_MAXCHILDREN] ;
 319   rsl_processor_t P[RSL_MAXCHILDREN] ;
 320 } rsl_child_info_t ;
 321
 322 typedef struct rsl_point {
 323   rsl_processor_t   P ;         /* physical processor number */
 324   int               valid ;     /* a valid point? */
 325   rsl_tag_t         trimmed ;   /* was this point trimmed */
 326   rsl_tag_t         cross ;     /* member of the cross grid */
 327   rsl_tag_t         info_1 ;    /* misc tag info */
 328   rsl_tag_t         info_2 ;    /* misc tag info */
 329   rsl_point_id_t    id ;
 330   rsl_point_id_t    mother_id ;
 331   rsl_processor_t   mother_P ;
 332   rsl_tag_t         which_kid_am_i_m ;
 333   rsl_tag_t         which_kid_am_i_n ;
 334
 335 /* on whole grid */
 336
 337                         /* counter-clockwise and clockwise below refer
 338                            to the direction of the traversal in fill_boundary
 339                            that determined the boundary association...
 340                            we store both to allow for several resolution
 341                            strategies on corners.  */
 342
 343 /* dot grid */
 344   rsl_index_t       bdy_cclockwise ;    /* direction to boundary (counter) */
 345   rsl_index_t       bdy_clockwise ;     /* direction to boundary (clockwise) */
 346   rsl_index_t       dbdy ;              /* distance to boundary */
 347
 348                         /* these next four fields are computed differently,
 349                            in that they do not take into account corners.
 350                            They simply measure the distance from the point
 351                            to the boundary in question. */
 352   rsl_index_t       dist_mlow ;         /* distance to southern bdy */
 353   rsl_index_t       dist_mhigh ;        /*     "    to northern bdy */
 354   rsl_index_t       dist_nlow ;         /*     "    to western bdy  */
 355   rsl_index_t       dist_nhigh ;        /*     "    to eastern bdy  */
 356
 357 /* on cross grid */
 358 /* cross grid */
 359   rsl_index_t       bdy_x_cclockwise ;  /* direction to boundary (counter) */
 360   rsl_index_t       bdy_x_clockwise ;   /* direction to boundary (clockwise) */
 361   rsl_index_t       dbdy_x ;            /* distance to boundary */
 362
 363   rsl_index_t       dist_mlow_x ;       /* distance to southern bdy */
 364   rsl_index_t       dist_mhigh_x ;      /*     "    to northern bdy */
 365   rsl_index_t       dist_nlow_x ;       /*     "    to western bdy  */
 366   rsl_index_t       dist_nhigh_x ;      /*     "    to eastern bdy  */
 367
 368   rsl_child_info_t *children_p ;
 369 } rsl_point_t ;
 370
 371 typedef struct rsl_hemi_rec {
 372   int oig , ojg ;
 373   int nbytes ;
 374   int curs ;
 375   char * data ;
 376   struct rsl_hemi_rec * next ;
 377 } rsl_hemi_rec_t ;
 378
 379 typedef struct rsl_domain_info {
 380   int valid ;
 381   int decomposed ;
 382
 383   /* parent domain descriptor */
 384   int parent ;
 385   int nestshape ;       /* RSL_REGULAR_NEST, RSL_RAGGED_NEST */
 386
 387   /* information about parent, children */
 388   int parent_bcast_compiled ;
 389   int parent_merge_compiled ;
 390   int child_bcast_compiled[RSL_MAXDOMAINS] ;
 391   int child_merge_compiled[RSL_MAXDOMAINS] ;
 392
 393   /* pointer to MN domain data structure (array of points) */
 394   rsl_point_t     *domain ;
 395
 396   /* pointer to MZ domain data structure (array of points) (20010222) */
 397   rsl_point_t     *domain_mz ;
 398
 399   /* pointer to NZ domain data structure (array of points) (20010222) */
 400   rsl_point_t     *domain_nz ;
 401
 402 /* MN decomp */
 403   /* dimensions of the global domain data structure */
 404   rsl_dimlen_t    len_m ;
 405   rsl_dimlen_t    len_n ;
 406   rsl_dimlen_t    eff_m ;
 407   rsl_dimlen_t    eff_n ;
 408   rsl_dimlen_t    loc_m ;
 409   rsl_dimlen_t    loc_n ;
 410   /* dimensions of the global domain data structure (20010222) */
 411   rsl_dimlen_t    len_z ;
 412   rsl_dimlen_t    eff_z ;
 413   rsl_dimlen_t    loc_z ;
 414 /* MZ decomp (20010223) */
 415   /* dimensions of the global domain data structure */
 416   rsl_dimlen_t    len_mz_m ;
 417   rsl_dimlen_t    len_mz_n ;
 418   rsl_dimlen_t    eff_mz_m ;
 419   rsl_dimlen_t    eff_mz_n ;
 420   rsl_dimlen_t    loc_mz_m ;
 421   rsl_dimlen_t    loc_mz_n ;
 422   rsl_dimlen_t    len_mz_z ;
 423   rsl_dimlen_t    eff_mz_z ;
 424   rsl_dimlen_t    loc_mz_z ;
 425 /* NZ decomp  (20010223 */
 426   /* dimensions of the global domain data structure */
 427   rsl_dimlen_t    len_nz_m ;
 428   rsl_dimlen_t    len_nz_n ;
 429   rsl_dimlen_t    eff_nz_m ;
 430   rsl_dimlen_t    eff_nz_n ;
 431   rsl_dimlen_t    loc_nz_m ;
 432   rsl_dimlen_t    loc_nz_n ;
 433   rsl_dimlen_t    len_nz_z ;
 434   rsl_dimlen_t    eff_nz_z ;
 435   rsl_dimlen_t    loc_nz_z ;
 436
 437   /* list of stencils used on this domain  (just keep handles) */
 438   int stenlist[RSL_MAXDESCRIPTORS] ;
 439   int stencurs ;
 440   int periodlist[RSL_MAXDESCRIPTORS] ;
 441   int periodcurs ;
 442   int xposelist[RSL_MAXDESCRIPTORS] ;
 443   int xposecurs ;
 444
 445   message_desc_t *old_state_vect, *new_state_vect ;
 446
 447   /* coordinates of the domain in its parent (sort of) */
 448   rsl_index_t     coord_m ;
 449   rsl_index_t     coord_n ;
 450
 451   /* nesting ratio in parent domain */
 452   int irax_m ;
 453   int irax_n ;
 454
 455   /* trimming constants -- these are used to tell rsl how much shorter
 456      in a dimension the effective range of a nest is than the declared
 457      range -- it might be declared longer because it has to be a
 458      multiple of irax_[mn] */
 459   int trim_m ;
 460   int trim_n ;
 461
 462   rsl_index_t  nest_level ;     /* 0 is mother domain */
 463   rsl_list_t   *pts ;           /* list of points on this processor */
 464   rsl_list_t   *ghost_pts ;     /* list of points communicating with us */
 465   rsl_list_t   *iruns ;         /* list of runs in idimension 1/9/95 */
 466   rsl_index_t  ilocaloffset ;   /* ns offset global to local in domain ds */
 467   rsl_index_t  jlocaloffset ;   /* ew offset global to local in domain ds */
 468
 469   rsl_index_t  ilocaloffset_mz ;
 470   rsl_index_t  jlocaloffset_mz ;
 471   rsl_index_t  klocaloffset_mz ;
 472
 473   rsl_index_t  ilocaloffset_nz ;
 474   rsl_index_t  jlocaloffset_nz ;
 475   rsl_index_t  klocaloffset_nz ;
 476
 477   rsl_index_t  old_ilocaloffset ;
 478   rsl_index_t  old_jlocaloffset ;
 479   int          maskid  ;        /* maximum stencil specified for this domain */
 480
 481 /* added for RSL_COMPUTE -- in comp_world.c */
 482 /* these are worked out in rsl_new_decomp.c */
 483   int    idif, jdif ;
 484   /* these are for iterating with j-major */
 485   int    nrun[MAX_KINDPAD+1];
 486   int    *(js[MAX_KINDPAD+1]) ;
 487   int    *(is[MAX_KINDPAD+1]), *(ie[MAX_KINDPAD+1]), *(jg2n[MAX_KINDPAD+1]) ;
 488   /* these are for iterating with i-major */
 489   int    nruni[MAX_KINDPAD+1];
 490   int    *(is2[MAX_KINDPAD+1]) ;
 491   int    *(js2[MAX_KINDPAD+1]), *(je2[MAX_KINDPAD+1]), *(ig2n[MAX_KINDPAD+1]) ;
 492
 493   rsl_processor_t bcast_recv_Pnpts[RSL_MAXPROC] ;
 494   rsl_processor_t bcast_recv_Plist[RSL_MAXPROC] ;
 495   rsl_processor_t bcast_recv_Ptags[RSL_MAXPROC] ;
 496   int Nbcast_recv_Plist ;
 497   rsl_processor_t bcast_send_Pnpts[RSL_MAXPROC] ;
 498   rsl_processor_t bcast_send_Plist[RSL_MAXPROC] ;
 499   int Nbcast_send_Plist ;
 500
 501   rsl_list_t *bcast_Xlist ;
 502
 503
 504   rsl_processor_t merge_recv_Pnpts[RSL_MAXPROC] ;
 505   rsl_processor_t merge_recv_Plist[RSL_MAXPROC] ;
 506   rsl_processor_t merge_recv_Ptags[RSL_MAXPROC] ;
 507   int Nmerge_recv_Plist ;
 508   rsl_processor_t merge_send_Pnpts[RSL_MAXPROC] ;
 509   rsl_processor_t merge_send_Plist[RSL_MAXPROC] ;
 510   int Nmerge_send_Plist ;
 511
 512   rsl_list_t *merge_Xlist ;
 513
 514
 515   int other_hemi_proclist_built ;
 516   rsl_hemi_rec_t * other_hemi_procbufs[RSL_MAXPROC] ;
 517   int hemi_sendPlist[RSL_MAXPROC] ;
 518   int hemi_recvPlist[RSL_MAXPROC] ;
 519   int hemi_recv_tags[RSL_MAXPROC] ;
 520
 521   int is_write, is_read, ie_write, ie_read ;
 522   int js_write, js_read, je_write, je_read ;
 523
 524 } rsl_domain_info_t ;
 525
 526
 527 /* March 1998 -- structure for new packing strategy */
 528 #ifdef crayx1
 529 struct packrec_struct {
 530     void * base ;
 531     int offset ;
 532     int n ;             /* number of bytes in an element */
 533     int nelems ;        /* number of elements */
 534     int stride ;        /* number of bytes between each element */
 535     int f90_table_index ;
 536     int endstop ;
 537     int valid ;
 538     int curs ;          /* position of data in pack buf */
 539 } ;
 540 typedef struct packrec_struct packrec_t ;
 541
 542 #else
 543 struct packrec_struct {
 544   void * base ;
 545   int offset ;
 546   int n ;               /* number of bytes in an element */
 547   int nelems ;          /* number of elements */
 548   int stride ;          /* number of bytes between each element */
 549   int f90_table_index ;
 550   int endstop ;
 551   int valid ;
 552 } ;
 553 typedef struct packrec_struct packrec_t ;
 554 #endif
 555
 556 typedef struct rsl_procrec {
 557   struct rsl_procrec * next ;
 558   rsl_processor_t P ;
 559   int sendsize ;        /* size of send buf needed for this processor */
 560   int recvsize ;        /* size of recv buf needed for this processor */
 561   int npts ;
 562   int recv_npts ;
 563   int nsends ;          /* diagnostic -- keeps running total number of sends */
 564   int nrecvs ;          /* diagnostic -- keeps running total number of sends */
 565   rsl_list_t *point_list ;
 566 /* Mar 1998 */
 567   rsl_list_t *recv_point_list ;
 568   packrec_t * pack_table ;
 569   int pack_table_size ;
 570   int pack_table_nbytes ;
 571   packrec_t * unpack_table ;
 572   int unpack_table_size ;
 573   int unpack_table_nbytes ;
 574 } rsl_procrec_t ;
 575
 576 typedef struct rsl_ptrec {
 577   struct rsl_ptrec * next ;
 578   rsl_point_t *pt ;
 579   rsl_index_t ig, jg ;
 580   rsl_list_t  *send_messages ;
 581   rsl_list_t  *recv_messages ;
 582 #ifdef crayx1
 583   int nsendmsgs ;
 584   int nrecvmsgs ;
 585 #else
 586   short nsendmsgs ;
 587   short nrecvmsgs ;
 588 #endif
 589 } rsl_ptrec_t ;
 590
 591 typedef struct rsl_point_hdr {
 592   rsl_index_t d, ig, jg, dummy ;   /* dummy maintains alignment */
 593   int nmsgs ;
 594 } rsl_point_hdr_t ;
 595
 596 typedef struct rsl_message_hdr {  /* packet hdr for message */
 597   int sp ;
 598 } rsl_message_hdr_t ;
 599
 600 typedef struct rsl_fld_hdr {    /* packet hdr for fld */
 601   void * base ;                 /* base address; byte pointer arithmetic */
 602 #ifdef crayx1
 603   int len ;                     /* number of bytes to follow */
 604 #else
 605   short len ;                   /* number of bytes to follow */
 606 #endif
 607 } rsl_fld_hdr_t ;
 608
 609 /* global data */
 610
 611 #ifdef DEFINE_GLOBAL
 612 #   define EXTERN
 613 #else
 614 #   define EXTERN extern
 615 #endif
 616
 617 EXTERN rsl_domain_info_t  domain_info[RSL_MAXDOMAINS] ;
 618 /* message descriptors */
 619 EXTERN void * mh_descriptors[RSL_MAXDESCRIPTORS] ;
 620 /* stencil descriptors */
 621 EXTERN void * sh_descriptors[RSL_MAXDESCRIPTORS] ;
 622 /* xpose descriptors */
 623 EXTERN void * xp_descriptors[RSL_MAXDESCRIPTORS] ;
 624 /* period descriptors */
 625 EXTERN void * pr_descriptors[RSL_MAXDESCRIPTORS] ;
 626 EXTERN int rsl_ndomains  ; /* number of active domains */
 627 EXTERN int rsl_nproc     ; /* number of compute processors */
 628 EXTERN int rsl_nproc_all ; /* total # of processors (compute and monitor) */
 629 EXTERN int rsl_nproc_n ; /* number of processors decomposing maj dimension */
 630 EXTERN int rsl_nproc_m ; /* number of processors decomposing min dimension */
 631 EXTERN int rsl_myproc ; /* my physical processor id */
 632 EXTERN int rsl_idum ;   /* dummy integer variable */
 633 EXTERN int rsl_padarea ;   /* pad area */
 634 EXTERN int io_seq_compute ;
 635 EXTERN int io_seq_monitor ;
 636 EXTERN char mess[1024] ;      /* misc. message buffer for errors and warns */
 637 EXTERN char * rsl_noprobe ;  /* set from environment */
 638
 639 EXTERN int rsl_debug_flg ;      /* set by rsl_debug */
 640 EXTERN int old_offsets ;        /* used in rsl_new_decomp.c */
 641
 642 EXTERN int regular_decomp ;
 643 EXTERN int sw_allow_dynpad ;
 644
 645 /* rsl macros */
 646
 647 #define RSL_FATAL(N)     rsl_fatal(N)
 648 #define RSL_TEST_ERR(T,M) {if(T){fprintf(stderr,"%d rsl error (\"%s\":%d) %s\n",rsl_myproc,__FILE__,__LINE__,M);RSL_FATAL(5);}}
 649 #define RSL_TEST_WRN(T,M) {if(T){fprintf(stderr,"%d rsl warning (\"%s\":%d) %s\n",rsl_myproc,__FILE__,__LINE__,M);}}
 650
 651 #if 1
 652 #ifndef NEC_TUNE
 653 #define RSL_MALLOC(T,N)  (T *)rsl_malloc(__FILE__,__LINE__,(sizeof(T))*(N))
 654 #define RSL_FREE(P)      rsl_free(P)
 655 #else
 656 #define RSL_MALLOC(T,N)  (T *)calloc(N, sizeof(T))
 657 #define RSL_FREE(P)      free(P)
 658 #endif
 659 #else
 660 /* Bob Olson's stuff */
 661 #define RSL_MALLOC(T,N)  (T *)nexus_debug_malloc((sizeof(T))*(N),__FILE__,__LINE__)
 662 #define RSL_FREE(P)      nexus_debug_free(P,__FILE__,__LINE__)
 663 #endif
 664
 665 #define MONITOR_ONLY(S)  { int x ; \
 666     RSL_C_IAMMONITOR( &x ) ; \
 667     if ( x==0 )                 \
 668       sprintf(mess,"%s: callable only by monitor",S) ; \
 669       RSL_TEST_ERR( x == 0,mess) ; }
 670
 671 /*
 672    This uses the low 26 bits of a 32 bit message type to encode a message
 673    type, a FROM processor designation and a TO processor designation.  Note
 674    the limits of this encoding:
 675
 676         o  10 bits available for processor id (limit is 1024 processors)
 677         o   6 bits available for message tag - 1, because we are not
 678             allowing 0.  (limit is thus 63 possible tags)
 679
 680    This is a potential machine or message passing interface dependency.
 681    Note, though, that under the proposed MPI standard, this would not
 682    be a problem because there are no FORCED type ranges to avoid and
 683    all .
 684
 685 rev: 95/02/15 -- jm.  Modified to drop the TO field (never used) and
 686    shorten the TAG field by 1 bit so that the total number of bits
 687    required is 15 (compatibility with minimum MPI standard and, in
 688    particular for the Fujitsu AP1000 at ANU).  Note the new limits
 689    from this encoding:
 690
 691         o  10 bits available for processor id (limit is 1024 processors)
 692            (FROM only is encoded now)
 693         o   5 bits available for message tag - 1, because we are not
 694             allowing 0.  (limit is now 31 possible tags).
 695 */
 696
 697 #if 0
 698 # define MTYPE_FROMTO(Y,F,T) ((((Y)&0x3f)<<20)|(((F)&0x3ff)<<10)|(((T)&0x3ff)))
 699 # define MTYPE_TAG(X)   (((X)>>20)&0x3f)
 700 # define MTYPE_FROM(X)  (((X)>>10)&0x3ff)
 701 # define MTYPE_TO(X)    ((X)&0x3ff)
 702 #else
 703 # if 1
 704 /* T is ignored and not encoded in the message tag, Y has been shorted 1 bit */
 705 # define MTYPE_FROMTO(Y,F,T) (((((Y)&0x1f)<<10)|(((F)&0x3ff))) | 0 )
 706 # define MTYPE_TAG(X)   (((X)>>10)&0x1f)
 707 # define MTYPE_FROM(X)  ((X)&0x3ff)
 708 # else
 709 /* update May 2002, increase upper bound on processors to 4096 (12 bits)*/
 710 # define MTYPE_FROMTO(Y,F,T) (((((Y)&0x1f)<<12)|(((F)&0xfff))) | 0 )
 711 # define MTYPE_TAG(X)   (((X)>>12)&0x1f)
 712 # define MTYPE_FROM(X)  ((X)&0xfff)
 713 # endif
 714 #endif
 715
 716
 717 #define INDEX_2(A,B,NB)       ( (B) + (A)*(NB) )
 718 #define INDEX_3(A,B,NB,C,NC)  INDEX_2( (A), INDEX_2( (B), (C), (NC) ), (NB)*(NC) )
 719
 720 #if 0
 721 /* new encoding for these --
 722     7 bits -- domain id         maximum number is 127
 723    12 bits -- ig index          maximum number is 4095
 724    12 bits -- jg index          maximum number is 4095
 725   ----
 726    31 bits
 727 */
 728 #define POINTID(D,J,I)   ((((((D)&0x7f))<<24)|(((J)&0xfff)<<12)|((I)&0xfff))|0L)
 729 #define ID_DOMAIN(M)     ((((M)>>24)&0x7f)|0L)
 730 #define ID_JDEX(M)       ((((M)>>12)&0xfff)|0L)
 731 #define ID_IDEX(M)       (((M)      &0xfff)|0L)
 732 #else
 733 /* make ig and jg work as signed quantities */
 734 /* new encoding for these --
 735     7 bits -- domain id         maximum number is 127
 736    12 bits -- ig index          range is -2048 to 2047
 737    12 bits -- jg index          range is is -2048 to 2047
 738   ----
 739    31 bits
 740 */
 741 #define POINTID(D,J,I)   ((((((D)&0x7f))<<24)|(((J)&0xfff)<<12)|((I)&0xfff))|0L)
 742 #define ID_DOMAIN(M)     ((((M)>>24)&0x7f)|0L)
 743 #define ID_JDEX(M)       (((((M)>>12)&0xfff)&0x800)?((((M)>>12)&0xfff)|(-4096L)):((((M)>>12)&0xfff)|0L))
 744 #define ID_IDEX(M)       (((M)&0x800)?(((M)      &0xfff)|(-4096L)):(((M)      &0xfff)|0L))
 745 #endif
 746
 747 #define IS_INVALID(M) ( (M) == RSL_INVALID )
 748
 749 /* forward declarations */
 750
 751 rsl_point_id_t pointid() ;
 752 rsl_index_t id_domain(), id_jdex(), id_idex() ;
 753
 754 /* RSL PT DESCRIPTORS */
 755
 756 #define RSL_4PT     4           /* FORTRAN */
 757 #define RSL_8PT     8           /* FORTRAN */
 758 #define RSL_12PT   12           /* FORTRAN */
 759 #define RSL_24PT   24           /* FORTRAN */
 760 #define RSL_48PT   48           /* FORTRAN */
 761 #define RSL_80PT   80           /* FORTRAN */
 762 #define RSL_120PT  120          /* FORTRAN */
 763 #if (ALLOW_RSL_168PT == 1)
 764 /* new -- added for MM5's upper radiative boundary stencil 9/26/96 */
 765 #define RSL_168PT  168          /* FORTRAN */
 766 #endif
 767
 768 #define RSL_NORTHSOUTH       1  /* FORTRAN */
 769 #define RSL_EASTWEST         2  /* FORTRAN */
 770 #define RSL_NORTHSOUTH_STAG  11 /* FORTRAN */
 771 #define RSL_EASTWEST_STAG    12 /* FORTRAN */
 772 #define RSL_NOTDECOMPOSED    3  /* FORTRAN */
 773
 774 #define RSL_M                1  /* FORTRAN */
 775 #define RSL_N                2  /* FORTRAN */
 776 #define RSL_M_STAG          11  /* FORTRAN */
 777 #define RSL_N_STAG          12  /* FORTRAN */
 778
 779 /* message tags */
 780
 781 #define MSG_NEWDECOMPOSITION 100
 782 #define MSG_MONITOR_REQUEST  101
 783 #define MSG_STENCOM                     1
 784 #define MSG_READ_RESPONSE               2
 785 #define MSG_WRITE_RESPONSE              3
 786 #define MSG_WRITE_COMPUTE_RESPONSE      4
 787 #define MSG_SPECIAL1_RESPONSE           5
 788 #define MSG_SPECIAL2_RESPONSE           6
 789 #define MSG_FROM_PARENT                 7
 790 #define MSG_BCAST_SETUP                 8
 791 #define MSG_MERGE_SETUP                 9
 792 #define MSG_TO_PARENT                  10
 793 #define MSG_MON_BCAST                  11
 794 #define MSG_REDISTCOM                  12
 795 #define MSG_PERCOM                     13
 796 #define MSG_XPOSECOM                     14
 797
 798 #define MSG_IO_FORTRAN                  1
 799 #define MSG_IO_SOCKET                   2
 800
 801 /* values for request mode2 when writing to sockets
 802    Raw mode just writes a byte stream,
 803
 804    Fortran mode puts control words at beginning and end of each write
 805    with byte count in control word.
 806
 807    Portal mode puts a portal style dimension descriptor at the
 808    head of each new write.
 809 */
 810
 811 #define MSG_MODE2_RAW                         1
 812 #define MSG_MODE2_FORTRAN                     2
 813 #define MSG_MODE2_PORTAL                      3
 814 #define MSG_MODE2_88                          4
 815
 816 /* monitor request types */
 817 #define RSL_READ_REQUEST      100
 818 #define RSL_READ_RESPONSE     101
 819 #define RSL_READ_SPECIAL1     102
 820 #define RSL_READ_SPECIAL2     103
 821 #define RSL_READ_SPECIAL3     104
 822 #define RSL_SHUTDOWN_REQUEST  105
 823 #define RSL_WRITE_REQUEST     106
 824 #define RSL_WRITE_RESPONSE    107
 825
 826 /* descriptor tags */
 827
 828 #define MESSAGE_DESC 0
 829 #define STENCIL_DESC 1
 830 #define BLANK_MESSAGE_DESC 2
 831 #define PERIOD_DESC 3
 832 #define XPOSE_DESC 4
 833
 834 /* xpose switches */
 835 #define XPOSE_MN_MZ 0
 836 #define XPOSE_MZ_MN 0
 837
 838 #define XPOSE_MZ_NZ 1
 839 #define XPOSE_NZ_MZ 1
 840
 841 #define XPOSE_MN_NZ 2
 842 #define XPOSE_NZ_MN 2
 843
 844 /* defines for MPI2 compat */
 845
 846 #ifndef MPI2_SUPPORT
 847 typedef int MPI_Fint;
 848 #  define MPI_Comm_c2f(comm) (MPI_Fint)(comm)
 849 #  define MPI_Comm_f2c(comm) (MPI_Comm)(comm)
 850 #endif
 851
 852 /* other includes */
 853
 854 #include "rsl_comm.h"
 855 #include "compat.h"
 856 #include "buf_for_proc.h"
 857 #include "stencil_def.h"        /* stencil_def.h must follow message_def.h */
 858 #include "xpose_def.h"  /* xpose_def.h must follow message_def.h */
 859 #include "period_def.h" /* period_def.h must follow message_def.h */
 860 #include "rsl_io.h"
 861
 862 #if (( defined(vpp) || defined(vpp2) ) && ! defined(sx))
 863 #define bcopy(a,b,c) vbcopy_C(a,b,c)
 864 #endif
 865
 866 void * rsl_malloc(), * malloc()  ;
 867 void * get_base_for_index() ;
 868
 869 #ifdef NEC_TUNE
 870 extern void copymem( void *, int, void *, int, int, int ) ;
 871 #endif
 872
 873
 874 #endif   /* nothing after this line */
 875
 876