TBCI Numerical high perf. C++ Library: ssp

00001 
00006 /* $Id: ssp_defs.h,v 1.3 2000/08/19 23:32:38 garloff Exp $ */
00007 
00008 /*
00009  * -- SuperLU routine (version 1.1) --
00010  * Univ. of California Berkeley, Xerox Palo Alto Research Center,
00011  * and Lawrence Berkeley National Lab.
00012  * November 15, 1997
00013  *
00014  */
00015 #ifndef __SUPERLU_sSP_DEFS /* allow multiple inclusions */
00016 #define __SUPERLU_sSP_DEFS
00017 
00018 /*
00019  * File name:           ssp_defs.h
00020  * Purpose:             Sparse matrix types and function prototypes
00021  * History:
00022  */
00023 #ifdef _CRAY
00024 #include <fortran.h>
00025 #include <string.h>
00026 #endif
00027 #include "Cnames.h"
00028 #include "supermatrix.h"
00029 
00030 
00031 /* No of marker arrays used in the symbolic factorization,
00032    each of size n */
00033 #define NO_MARKER     3
00034 #define NUM_TEMPV(m,w,t,b)  ( MAX(m, (t + b)*w) )
00035 
00036 typedef enum {LUSUP, UCOL, LSUB, USUB} MemType;
00037 typedef enum {HEAD, TAIL}              stack_end_t;
00038 typedef enum {SYSTEM, USER}            LU_space_t;
00039 
00040 /*
00041  * Global data structures used in LU factorization -
00042  * 
00043  *   nsuper: #supernodes = nsuper + 1, numbered [0, nsuper].
00044  *   (xsup,supno): supno[i] is the supernode no to which i belongs;
00045  *      xsup(s) points to the beginning of the s-th supernode.
00046  *      e.g.   supno 0 1 2 2 3 3 3 4 4 4 4 4   (n=12)
00047  *              xsup 0 1 2 4 7 12
00048  *      Note: dfs will be performed on supernode rep. relative to the new 
00049  *            row pivoting ordering
00050  *
00051  *   (xlsub,lsub): lsub[*] contains the compressed subscript of
00052  *      rectangular supernodes; xlsub[j] points to the starting
00053  *      location of the j-th column in lsub[*]. Note that xlsub 
00054  *      is indexed by column.
00055  *      Storage: original row subscripts
00056  *
00057  *      During the course of sparse LU factorization, we also use
00058  *      (xlsub,lsub) for the purpose of symmetric pruning. For each
00059  *      supernode {s,s+1,...,t=s+r} with first column s and last
00060  *      column t, the subscript set
00061  *              lsub[j], j=xlsub[s], .., xlsub[s+1]-1
00062  *      is the structure of column s (i.e. structure of this supernode).
00063  *      It is used for the storage of numerical values.
00064  *      Furthermore,
00065  *              lsub[j], j=xlsub[t], .., xlsub[t+1]-1
00066  *      is the structure of the last column t of this supernode.
00067  *      It is for the purpose of symmetric pruning. Therefore, the
00068  *      structural subscripts can be rearranged without making physical
00069  *      interchanges among the numerical values.
00070  *
00071  *      However, if the supernode has only one column, then we
00072  *      only keep one set of subscripts. For any subscript interchange
00073  *      performed, similar interchange must be done on the numerical
00074  *      values.
00075  *
00076  *      The last column structures (for pruning) will be removed
00077  *      after the numercial LU factorization phase.
00078  *
00079  *   (xlusup,lusup): lusup[*] contains the numerical values of the
00080  *      rectangular supernodes; xlusup[j] points to the starting
00081  *      location of the j-th column in storage vector lusup[*]
00082  *      Note: xlusup is indexed by column.
00083  *      Each rectangular supernode is stored by column-major
00084  *      scheme, consistent with Fortran 2-dim array storage.
00085  *
00086  *   (xusub,ucol,usub): ucol[*] stores the numerical values of
00087  *      U-columns outside the rectangular supernodes. The row
00088  *      subscript of nonzero ucol[k] is stored in usub[k].
00089  *      xusub[i] points to the starting location of column i in ucol.
00090  *      Storage: new row subscripts; that is subscripts of PA.
00091  */
00092 typedef struct {
00093     int     *xsup;    /* supernode and column mapping */
00094     int     *supno;   
00095     int     *lsub;    /* compressed L subscripts */
00096     int     *xlsub;
00097     float  *lusup;   /* L supernodes */
00098     int     *xlusup;
00099     float  *ucol;    /* U columns */
00100     int     *usub;
00101     int     *xusub;
00102     int     nzlmax;   /* current max size of lsub */
00103     int     nzumax;   /*    "    "    "      ucol */
00104     int     nzlumax;  /*    "    "    "     lusup */
00105     int     n;        /* number of columns in the matrix */
00106     LU_space_t MemModel; /* 0 - system malloc'd; 1 - user provided */
00107 } GlobalLU_t;
00108 
00109 typedef struct {
00110     int panel_size;
00111     int relax;
00112     float diag_pivot_thresh;
00113     float drop_tol;
00114 } factor_param_t;
00115 
00116 typedef struct {
00117     float for_lu;
00118     float total_needed;
00119     int   expansions;
00120 } mem_usage_t;
00121 
00122 #ifdef __cplusplus
00123 extern "C" {
00124 #endif
00125 
00126 /* Driver routines */
00127 extern void
00128 sgssv(SuperMatrix *, int *, int *, SuperMatrix *, SuperMatrix *, 
00129         SuperMatrix *, int *);
00130 extern void
00131 sgssvx(char *, char *, char *, SuperMatrix *, factor_param_t *,
00132        int *, int *, int *, char *, float *, float *,
00133        SuperMatrix *, SuperMatrix *, void *, int, SuperMatrix *, 
00134        SuperMatrix *, float *, float *, float *,
00135        float *, mem_usage_t *, int *);
00136 
00137 /* Supernodal LU factor related */
00138 extern void
00139 sCreate_CompCol_Matrix(SuperMatrix *, int, int, int, float *,
00140                        int *, int *, Stype_t, Dtype_t, Mtype_t);
00141 extern void
00142 sCopy_CompCol_Matrix(SuperMatrix *, SuperMatrix *);
00143 extern void
00144 sCreate_Dense_Matrix(SuperMatrix *, int, int, float *, int,
00145                      Stype_t, Dtype_t, Mtype_t);
00146 extern void
00147 sCreate_SuperNode_Matrix(SuperMatrix *, int, int, int, float *, 
00148                          int *, int *, int *, int *, int *,
00149                          Stype_t, Dtype_t, Mtype_t);
00150 extern void
00151 sCopy_Dense_Matrix(int, int, float *, int, float *, int);
00152 
00153 extern void Destroy_SuperMatrix_Store(SuperMatrix *);
00154 extern void Destroy_CompCol_Matrix(SuperMatrix *);
00155 extern void Destroy_SuperNode_Matrix(SuperMatrix *);
00156 extern void Destroy_CompCol_Permuted(SuperMatrix *);
00157 extern void Destroy_Dense_Matrix(SuperMatrix *);
00158 
00159 extern void    sallocateA (int, int, float **, int **, int **);
00160 extern void    StatInit    (int, int);
00161 extern void    StatFree    ();
00162 extern void    get_perm_c(int, SuperMatrix *, int *);
00163 extern void    sp_preorder (char*, SuperMatrix*, int*, int*, SuperMatrix*);
00164 extern int     sp_coletree (int *, int *, int *, int, int, int *);
00165 extern void    sgstrf (char*, SuperMatrix*, float, float, int, int, int*,
00166                         void *, int, int *, int *, 
00167                         SuperMatrix *, SuperMatrix *, int *);
00168 extern void    relax_snode  (int, int *, int, int *, int *);
00169 extern int     ssnode_dfs (const int, const int, const int *, const int *,
00170                              const int *, int *, int *, GlobalLU_t *);
00171 extern int     ssnode_bmod (const int, const int, const int, float *,
00172                               float *, GlobalLU_t *);
00173 extern void    spanel_dfs (const int, const int, const int, SuperMatrix *,
00174                            int *, int *, float *, int *, int *, int *,
00175                            int *, int *, int *, int *, GlobalLU_t *);
00176 extern void    spanel_bmod (const int, const int, const int, const int,
00177                            float *, float *, int *, int *,
00178                            GlobalLU_t *);
00179 extern int     scolumn_dfs (const int, const int, int *, int *, int *, int *,
00180                            int *, int *, int *, int *, int *, GlobalLU_t *);
00181 extern int     scolumn_bmod (const int, const int, float *,
00182                            float *, int *, int *, int, GlobalLU_t *);
00183 extern int     scopy_to_ucol (int, int, int *, int *, int *,
00184                               float *, GlobalLU_t *);         
00185 extern int     spivotL (const int, const float, int *, int *, 
00186                               int *, int *, int *, GlobalLU_t *);
00187 extern void    spruneL (const int, const int *, const int, const int,
00188                              const int *, const int *, int *, GlobalLU_t *);
00189 extern void    resetrep_col (const int, const int *, int *);
00190 extern void    countnz (const int, int *, int *, int *, GlobalLU_t *);
00191 extern void    fixupL (const int, const int *, GlobalLU_t *);
00192 extern int     spcoletree (int *, int *, int *, int, int, int *);
00193 extern int     *TreePostorder (int, int *);
00194 extern void    sreadmt (int *, int *, int *, float **, int **, int **);
00195 extern void    sGenXtrue (int, int, float *, int);
00196 extern void    sFillRHS (char *, int, float *, int, SuperMatrix *,
00197                         SuperMatrix *);
00198 extern void    sgstrs (char *, SuperMatrix *, SuperMatrix *, int *, int *,
00199                         SuperMatrix *, int *);
00200 
00201 
00202 /* Driver related */
00203 
00204 extern void    sgsequ (SuperMatrix *, float *, float *, float *,
00205                              float *, float *, int *);
00206 extern void    slaqgs (SuperMatrix *, float *, float *, float,
00207                              float, float, char *);
00208 extern void    sgscon (char *, SuperMatrix *, SuperMatrix *, 
00209                         float, float *, int *);
00210 extern float   sPivotGrowth(int, SuperMatrix *, int *, 
00211                             SuperMatrix *, SuperMatrix *);
00212 extern void    sgsrfs (char *, SuperMatrix *, SuperMatrix *,
00213                         SuperMatrix *, int *, int *, char *, float *, 
00214                         float *, SuperMatrix *, SuperMatrix *, float *,
00215                         float *, int *);
00216 
00217 extern int     sp_strsv (char *, char *, char *, SuperMatrix *,
00218                         SuperMatrix *, float *, int *);
00219 extern int     sp_sgemv (char *, float, SuperMatrix *, float *,
00220                         int, float, float *, int);
00221 
00222 extern int     sp_sgemm (char *, char *, int, int, int, float,
00223                         SuperMatrix *, float *, int, float, 
00224                         float *, int);
00225 
00226 /* Memory-related */
00227 extern void     superlu_abort_and_exit(char*);
00228 extern void    *superlu_malloc (int);
00229 extern void    superlu_free (void*);
00230 extern int     sLUMemInit (char *, void *, int, int, int, int, int,
00231                              SuperMatrix *, SuperMatrix *,
00232                              GlobalLU_t *, int **, float **);
00233 extern void    SetIWork (int, int, int, int *, int **, int **, int **,
00234                          int **, int **, int **, int **);
00235 extern void    sSetRWork (int, int, float *, float **, float **);
00236 extern void    sLUWorkFree (int *, float *, GlobalLU_t *);
00237 extern int     sLUMemXpand (int, int, MemType, int *, GlobalLU_t *);
00238 
00239 extern int     *intMalloc (int);
00240 extern int     *intCalloc (int);
00241 extern float  *floatMalloc(int);
00242 extern float  *floatCalloc(int);
00243 extern void    *superlu_malloc(int);
00244 extern int     memory_usage();
00245 extern int     sQuerySpace (SuperMatrix *, SuperMatrix *, int,
00246                                 mem_usage_t *);
00247 
00248 /* Auxiliary routines */
00249 extern void    sCompRow_to_CompCol(int, int, int, float*, int*, int*,
00250                                    float **, int **, int **);
00251 extern double  SuperLU_timer_ ();
00252 extern int     sp_ienv (int);
00253 extern int     lsame_ (char *, char *);
00254 extern int     xerbla_ (char *, int *);
00255 extern void    ifill (int *, int, int);
00256 extern void    sfill (float *, int, float);
00257 extern void    inf_norm_error (int, SuperMatrix *, float *);
00258 extern void    snode_profile (int, int *);
00259 extern void    super_stats (int, int *);
00260 extern void    PrintSumm (char *, int, int, int);
00261 extern void    PrintPerf (SuperMatrix *, SuperMatrix *, mem_usage_t *,
00262                          float, float, float *, float *, char *);
00263 
00264 /* Routines for debugging */
00265 extern void    sPrint_CompCol_Matrix(char *, SuperMatrix *);
00266 extern void    sPrint_SuperNode_Matrix(char *, SuperMatrix *);
00267 extern void    sPrint_Dense_Matrix(char *, SuperMatrix *);
00268 extern void    print_lu_col(char *, int, int, int *, GlobalLU_t *);
00269 extern void    print_panel_seg(int, int, int, int, int *, int *);
00270 extern void    check_tempv(int, float *);
00271 extern void    check_repfnz(int, int, int, int *);
00272 
00273 #ifdef __cplusplus
00274   }
00275 #endif
00276 
00277 #endif /* __SUPERLU_sSP_DEFS */
00278