File: distributed_qsort.c

package info (click to toggle)
hypre 2.8.0b-1
  • links: PTS, VCS
  • area: main
  • in suites: wheezy
  • size: 147,420 kB
  • sloc: ansic: 1,781,700; f90: 251,025; cpp: 218,266; fortran: 51,552; sh: 17,720; java: 12,225; makefile: 5,371; python: 1,096; awk: 147; pascal: 92; perl: 56; lisp: 28; csh: 11
file content (241 lines) | stat: -rw-r--r-- 6,377 bytes parent folder | download | duplicates (2)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
/*BHEADER**********************************************************************
 * Copyright (c) 2008,  Lawrence Livermore National Security, LLC.
 * Produced at the Lawrence Livermore National Laboratory.
 * This file is part of HYPRE.  See file COPYRIGHT for details.
 *
 * HYPRE is free software; you can redistribute it and/or modify it under the
 * terms of the GNU Lesser General Public License (as published by the Free
 * Software Foundation) version 2.1 dated February 1999.
 *
 * $Revision: 2.7 $
 ***********************************************************************EHEADER*/




/*
 * distributed_qsort.c:
 * Our own version of the system qsort routine which is faster by an average
 * of 25%, with lows and highs of 10% and 50%.
 * The THRESHold below is the insertion sort threshold, and has been adjusted
 * for records of size 48 bytes.
 * The MTHREShold is where we stop finding a better median.
 */

#include <stdlib.h>			/* only for type declarations */
#include "_hypre_utilities.h"

#define		THRESH		4	/* threshold for insertion */
#define		MTHRESH		6	/* threshold for median */

static HYPRE_Int (*qcmp) ();		/* the comparison routine */
static HYPRE_Int qsz;			/* size of each record */
static void qst(char *, char *);

static HYPRE_Int thresh;		/* THRESHold in chars */
static HYPRE_Int mthresh;		/* MTHRESHold in chars */



/*
 * hypre_tex_qsort:
 * First, set up some global parameters for qst to share.  Then, quicksort
 * with qst(), and then a cleanup insertion sort ourselves.  Sound simple?
 * It's not...
 */

void
hypre_tex_qsort(base, n, size, compar)
char *base;
HYPRE_Int n;
HYPRE_Int size;
HYPRE_Int (*compar) ();
{
    register char *i;
    register char *j;
    register char *lo;
    register char *hi;
    register char *min;
    register char c;
    char *max;

    if (n <= 1)
	return;
    qsz = size;
    qcmp = compar;
    thresh = qsz * THRESH;
    mthresh = qsz * MTHRESH;
    max = base + n * qsz;
    if (n >= THRESH)
    {
	qst(base, max);
	hi = base + thresh;
    }
    else
    {
	hi = max;
    }
    /* First put smallest element, which must be in the first THRESH, in the
       first position as a sentinel.  This is done just by searching the
       first THRESH elements (or the first n if n < THRESH), finding the min,
       and swapping it into the first position. */
    for (j = lo = base; (lo += qsz) < hi;)
    {
	if ((*qcmp) (j, lo) > 0)
	    j = lo;
    }
    if (j != base)
    {				/* swap j into place */
	for (i = base, hi = base + qsz; i < hi;)
	{
	    c = *j;
	    *j++ = *i;
	    *i++ = c;
	}
    }
    /* With our sentinel in place, we now run the following hyper-fast
       insertion sort.  For each remaining element, min, from [1] to [n-1],
       set hi to the index of the element AFTER which this one goes. Then, do
       the standard insertion sort shift on a character at a time basis for
       each element in the frob. */
    for (min = base; (hi = min += qsz) < max;)
    {
	while ((*qcmp) (hi -= qsz, min) > 0);
	if ((hi += qsz) != min)
	{
	    for (lo = min + qsz; --lo >= min;)
	    {
		c = *lo;
		for (i = j = lo; (j -= qsz) >= hi; i = j)
		    *i = *j;
		*i = c;
	    }
	}
    }
}



/*
 * qst:
 * Do a quicksort
 * First, find the median element, and put that one in the first place as the
 * discriminator.  (This "median" is just the median of the first, last and
 * middle elements).  (Using this median instead of the first element is a big
 * win).  Then, the usual partitioning/swapping, followed by moving the
 * discriminator into the right place.  Then, figure out the sizes of the two
 * partions, do the smaller one recursively and the larger one via a repeat of
 * this code.  Stopping when there are less than THRESH elements in a partition
 * and cleaning up with an insertion sort (in our caller) is a huge win.
 * All data swaps are done in-line, which is space-losing but time-saving.
 * (And there are only three places where this is done).
 */

static void qst(char *base, char *max)
{
    register char *i;
    register char *j;
    register char *jj;
    register char *mid;
    register HYPRE_Int ii;
    register char c;
    char *tmp;
    HYPRE_Int lo;
    HYPRE_Int hi;

    lo = max - base;		/* number of elements as chars */
    do
    {
	/* At the top here, lo is the number of characters of elements in the
	   current partition.  (Which should be max - base). Find the median
	   of the first, last, and middle element and make that the middle
	   element.  Set j to largest of first and middle.  If max is larger
	   than that guy, then it's that guy, else compare max with loser of
	   first and take larger.  Things are set up to prefer the middle,
	   then the first in case of ties. */
	mid = i = base + qsz * ((unsigned) (lo / qsz) >> 1);
	if (lo >= mthresh)
	{
	    j = ((*qcmp) ((jj = base), i) > 0 ? jj : i);
	    if ((*qcmp) (j, (tmp = max - qsz)) > 0)
	    {
		j = (j == jj ? i : jj);	/* switch to first loser */
		if ((*qcmp) (j, tmp) < 0)
		    j = tmp;
	    }
	    if (j != i)
	    {
		ii = qsz;
		do
		{
		    c = *i;
		    *i++ = *j;
		    *j++ = c;
		} while (--ii);
	    }
	}
	/* Semi-standard quicksort partitioning/swapping */
	for (i = base, j = max - qsz;;)
	{
	    while (i < mid && (*qcmp) (i, mid) <= 0)
		i += qsz;
	    while (j > mid)
	    {
		if ((*qcmp) (mid, j) <= 0)
		{
		    j -= qsz;
		    continue;
		}
		tmp = i + qsz;	/* value of i after swap */
		if (i == mid)
		{		/* j <-> mid, new mid is j */
		    mid = jj = j;
		}
		else
		{		/* i <-> j */
		    jj = j;
		    j -= qsz;
		}
		goto swap;
	    }
	    if (i == mid)
	    {
		break;
	    }
	    else
	    {			/* i <-> mid, new mid is i */
		jj = mid;
		tmp = mid = i;	/* value of i after swap */
		j -= qsz;
	    }
    swap:
	    ii = qsz;
	    do
	    {
		c = *i;
		*i++ = *jj;
		*jj++ = c;
	    } while (--ii);
	    i = tmp;
	}
	/* Look at sizes of the two partitions, do the smaller one first by
	   recursion, then do the larger one by making sure lo is its size,
	   base and max are update correctly, and branching back. But only
	   repeat (recursively or by branching) if the partition is of at
	   least size THRESH. */
	i = (j = mid) + qsz;
	if ((lo = j - base) <= (hi = max - i))
	{
	    if (lo >= thresh)
		qst(base, j);
	    base = i;
	    lo = hi;
	}
	else
	{
	    if (hi >= thresh)
		qst(i, max);
	    max = j;
	}
    } while (lo >= thresh);
}