1 |
/* |
2 |
* Model-based Reordering Routines |
3 |
* by Benjamin Andrew Allan |
4 |
* 6/22/96 |
5 |
* Version: $Revision: 1.5 $ |
6 |
* Version control file: $RCSfile: model_reorder.h,v $ |
7 |
* Date last modified: $Date: 1997/07/18 12:14:44 $ |
8 |
* Last modified by: $Author: mthomas $ |
9 |
* Copyright(C) 1996 Benjamin Andrew Allan |
10 |
* |
11 |
* This file is part of the ASCEND IV math programming system. |
12 |
* These functions are part of a new design for feeding |
13 |
* solvers from the ASCEND compiler. |
14 |
* |
15 |
* File to play MODEL-relation based reordering games. |
16 |
* We're starting with a basic RBBD implementation. |
17 |
* Assumptions: |
18 |
* - If the MODELs are hierarchical, then they have been indexed in a |
19 |
* tree bottom-up fashion. |
20 |
* - Input is a square block that needs tearing and not a rectangle. |
21 |
* |
22 |
* The ASCEND IV math programming system is free software; you can redistribute |
23 |
* it and/or modify it under the terms of the GNU General Public License as |
24 |
* published by the Free Software Foundation; either version 2 of the |
25 |
* License, or (at your option) any later version. |
26 |
* |
27 |
* The ASCEND IV math programming system is distributed in hope that it will be |
28 |
* useful, but WITHOUT ANY WARRANTY; without even the implied warranty of |
29 |
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |
30 |
* General Public License for more details. |
31 |
* |
32 |
* You should have received a copy of the GNU General Public License along with |
33 |
* the program; if not, write to the Free Software Foundation, Inc., 675 |
34 |
* Mass Ave, Cambridge, MA 02139 USA. Check the file named COPYING. |
35 |
* COPYING is found in ../compiler. |
36 |
*/ |
37 |
#ifndef _model_reorder_h_seen_ |
38 |
#define _model_reorder_h_seen_ |
39 |
/* requires #include "base.h" */ |
40 |
/* requires #include "slv_types.h" */ |
41 |
/* requires #include "mtx.h" */ |
42 |
|
43 |
#define MRDEBUG 0 |
44 |
/* if !=0, generate spew while working */ |
45 |
#define CUTOFFDEFAULT 1000 |
46 |
/* tuning parameter in need of investigation */ |
47 |
|
48 |
typedef struct mr_bisection_structure { |
49 |
/* data spaces */ |
50 |
slv_system_t slv; |
51 |
mtx_matrix_t mtx; |
52 |
int32 *local; /* MODEL indexed array of counts of relations */ |
53 |
char *active; /* in block flags for models */ |
54 |
int32 *tmpmodlist; /* a unique list of indices of models found so far */ |
55 |
int32 tmpmodused; /* currently used length of tmpmodlist */ |
56 |
int32 activelen; /* length of active,local and tmpmodlist arrays */ |
57 |
|
58 |
/* control parameters */ |
59 |
int32 cutoff; |
60 |
int32 maxtears; |
61 |
double maxtearfraction; |
62 |
|
63 |
/* miscellaneous statistics */ |
64 |
int32 ntears; |
65 |
int32 nblts; |
66 |
} mr_reorder_t; |
67 |
/************************************************************************\ |
68 |
The mr_reorder_t is a workspace structure containing most of the items |
69 |
needed for the recursive function mr_bisect_partition and any similar |
70 |
functions. mr_reorder_create and destroy are provided to set up |
71 |
a mr_reorder_t. |
72 |
- The mtx herein should correspond to the solvers_rel/var lists of the |
73 |
slv_system_t here. |
74 |
- The pointers local, tmpmodlist, and active should contain arrays of |
75 |
size activelen and these arrays should be (re)initialized to 0 |
76 |
for each call of mr_bisect_partition. |
77 |
- Activelen should be the number of models in the slv_system_t. In |
78 |
particular array[rel_model(rel)] should exist for all rel in the |
79 |
solvers_rel_list. |
80 |
- Cutoff is the smallest block the user wants us to try to tear. |
81 |
- Maxtears is the largest number of tears user wants to allow in any |
82 |
single block. IT IS IGNORED. |
83 |
- Maxtearfraction is the largest percentage of any single block the |
84 |
user wants us to try to tear. IT IS IGNORED. |
85 |
- Ntears is the number of tear columns we find. |
86 |
- Nblts is the number of BLT permutations we do. |
87 |
\************************************************************************/ |
88 |
|
89 |
extern mr_reorder_t *mr_reorder_create(slv_system_t, mtx_matrix_t, int32); |
90 |
/************************************************************************\ |
91 |
mrsys = mr_reorder_create(slvsys,mtx,nmodels); |
92 |
Returns a mr_reorder_t all set up with cutoff set to CUTOFFDEFAULT |
93 |
for a problem that has nmodels models in it. |
94 |
The arrays are initialized to 0. |
95 |
rel_model(rel) of any rel in slvsys should return a number 0..nmodels-1. |
96 |
Returns NULL if insufficient memory. |
97 |
\************************************************************************/ |
98 |
|
99 |
extern void mr_reorder_destroy(mr_reorder_t *); |
100 |
/************************************************************************\ |
101 |
mr_reorder_destroy(mrsys); |
102 |
Deallocate a mr_reorder_t *. |
103 |
\************************************************************************/ |
104 |
|
105 |
typedef int (MRBlockReorderF)(slv_system_t,mtx_matrix_t,mtx_region_t *); |
106 |
/************************************************************************\ |
107 |
This is a user supplied function that will be applied called |
108 |
with sys->slv,sys->mtx, and regions that fall below sys->cutoff |
109 |
in size. The user may do any sort of logic they care to, |
110 |
including ignoring reordering blocks below a certain size. It |
111 |
must be supplied, even if it does nothing. |
112 |
|
113 |
This function could be dumb and just apply SPK1-like things, or |
114 |
be very smart and reorder based on the number and type of models |
115 |
found in the region. This function should be the subject of some |
116 |
experimentation since mr_bisect_partition just identifies tears |
117 |
and subregions. |
118 |
|
119 |
This function is also the one called on blocks bigger than the |
120 |
cutoff size that however contain equations from exactly 1 MODEL. |
121 |
|
122 |
We don't (yet, anyway) check the return value. |
123 |
\************************************************************************/ |
124 |
|
125 |
extern int mr_bisect_partition(mr_reorder_t *, mtx_region_t *, int, |
126 |
MRBlockReorderF); |
127 |
/************************************************************************\ |
128 |
stat = mr_bisect_partition(sys, reg, top, rfunc); |
129 |
mr_reorder_t *sys; |
130 |
mtx_region_t *reg; |
131 |
int top; |
132 |
MRBlockReorderF rfunc; |
133 |
|
134 |
This function is recursive and produces a sys->mtx reordered to be |
135 |
recursive block bordered diagonal (RBBD). |
136 |
|
137 |
If top is 1, the region given in block should be (within itself) |
138 |
a strongly connected square block. If top is 0, we will do a BLT |
139 |
permutation on the region first assuming the region is square |
140 |
and has a full diagonal. The normal mode of externally calling |
141 |
this function should be with top==1 and we then recurse with |
142 |
top==0. |
143 |
|
144 |
One the outermost call, the relations in reg should have all |
145 |
their REL_TORN flag bits set to 0. On return from the outermost |
146 |
call, the columns identified as tears will have the REL_TORN |
147 |
flag bits set to 1. It turns out we don't use the REL_PARTITION |
148 |
flag. |
149 |
|
150 |
The mr_reorder_t * given must be filled in completely before |
151 |
calling this function. Anything amiss will cause us to return |
152 |
immediately without doing anything. Returns 0 if ok, > 0 if |
153 |
anything else wrong. Basically, the only thing that can go |
154 |
wrong is, if during one of our blt permutations, we run out |
155 |
of memory. Nonzero return value will be the number of times |
156 |
we encountered malloc failure. |
157 |
|
158 |
The values in reg may be messed with, so if the data in |
159 |
reg is needed after this function call, keep it elsewhere. |
160 |
\************************************************************************/ |
161 |
|
162 |
extern int mr_bisect_partition2(mr_reorder_t *, mtx_region_t *, int, |
163 |
MRBlockReorderF); |
164 |
/************************************************************************\ |
165 |
stat = mr_bisect_partition2(sys, reg, top, rfunc); |
166 |
mr_reorder_t *sys; |
167 |
mtx_region_t *reg; |
168 |
int top; |
169 |
MRBlockReorderF rfunc; |
170 |
|
171 |
This is the function most similar to the algorithm described/tested |
172 |
in the Abbott thesis. |
173 |
|
174 |
This function is recursive and produces a sys->mtx reordered to be |
175 |
recursive block bordered diagonal (RBBD) with borders roughly twice |
176 |
the size of those in mr_bisect_partition because of a simple idiocy |
177 |
in the way it locates tears. The relative performance of the two |
178 |
versions is not clear. |
179 |
|
180 |
If top is 1, the region given in block should be (within itself) |
181 |
a strongly connected square block. If top is 0, we will do a BLT |
182 |
permutation on the region first assuming the region is square |
183 |
and has a full diagonal. The normal mode of externally calling |
184 |
this function should be with top==1 and we then recurse with |
185 |
top==0. |
186 |
|
187 |
One the outermost call, the relations in reg should have all |
188 |
their REL_TORN flag bits set to 0. On return from the outermost |
189 |
call, the columns identified as tears will have the REL_TORN |
190 |
flag bits set to 1. It turns out we don't use the REL_PARTITION |
191 |
flag. |
192 |
|
193 |
The mr_reorder_t * given must be filled in completely before |
194 |
calling this function. Anything amiss will cause us to return |
195 |
immediately without doing anything. Returns 0 if ok, > 0 if |
196 |
anything else wrong. Basically, the only thing that can go |
197 |
wrong is, if during one of our blt permutations, we run out |
198 |
of memory. Nonzero return value will be the number of times |
199 |
we encountered malloc failure. |
200 |
|
201 |
The values in reg may be messed with, so if the data in |
202 |
reg is needed after this function call, keep it elsewhere. |
203 |
\************************************************************************/ |
204 |
#endif /* _model_reorder_h_seen_ */ |