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
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
|
<!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 3.2 Final//EN">
<!--Converted with LaTeX2HTML 98.2 beta6 (August 14th, 1998)
original version by: Nikos Drakos, CBLU, University of Leeds
* revised and updated by: Marcus Hennecke, Ross Moore, Herb Swan
* with significant contributions from:
Jens Lippmann, Marek Rouchal, Martin Wilck and others -->
<HTML>
<HEAD>
<TITLE>Factorizations for Solving Linear Equations</TITLE>
<META NAME="description" CONTENT="Factorizations for Solving Linear Equations">
<META NAME="keywords" CONTENT="lug_l2h">
<META NAME="resource-type" CONTENT="document">
<META NAME="distribution" CONTENT="global">
<META HTTP-EQUIV="Content-Type" CONTENT="text/html; charset=iso-8859-1">
<LINK REL="STYLESHEET" HREF="lug_l2h.css">
<LINK REL="next" HREF="node69.html">
<LINK REL="previous" HREF="node67.html">
<LINK REL="up" HREF="node67.html">
<LINK REL="next" HREF="node69.html">
</HEAD>
<BODY >
<!--Navigation Panel-->
<A NAME="tex2html5115"
HREF="node69.html">
<IMG WIDTH="37" HEIGHT="24" ALIGN="BOTTOM" BORDER="0" ALT="next"
SRC="next_motif.png"></A>
<A NAME="tex2html5109"
HREF="node67.html">
<IMG WIDTH="26" HEIGHT="24" ALIGN="BOTTOM" BORDER="0" ALT="up"
SRC="up_motif.png"></A>
<A NAME="tex2html5103"
HREF="node67.html">
<IMG WIDTH="63" HEIGHT="24" ALIGN="BOTTOM" BORDER="0" ALT="previous"
SRC="previous_motif.png"></A>
<A NAME="tex2html5111"
HREF="node1.html">
<IMG WIDTH="65" HEIGHT="24" ALIGN="BOTTOM" BORDER="0" ALT="contents"
SRC="contents_motif.png"></A>
<A NAME="tex2html5113"
HREF="node152.html">
<IMG WIDTH="43" HEIGHT="24" ALIGN="BOTTOM" BORDER="0" ALT="index"
SRC="index_motif.png"></A>
<BR>
<B> Next:</B> <A NAME="tex2html5116"
HREF="node69.html">QR Factorization</A>
<B> Up:</B> <A NAME="tex2html5110"
HREF="node67.html">Examples of Block Algorithms</A>
<B> Previous:</B> <A NAME="tex2html5104"
HREF="node67.html">Examples of Block Algorithms</A>
  <B> <A NAME="tex2html5112"
HREF="node1.html">Contents</A></B>
  <B> <A NAME="tex2html5114"
HREF="node152.html">Index</A></B>
<BR>
<BR>
<!--End of Navigation Panel-->
<H2><A NAME="SECTION03341000000000000000"></A>
<A NAME="subsecblocklineq"></A>
<BR>
Factorizations for Solving Linear Equations
</H2>
<P>
The well-known <B><I>LU</I></B> and Cholesky factorizations are the simplest block
algorithms to derive. No extra floating-point operations nor extra
working storage are required.
<P>
Table <A HREF="node68.html#tablu">3.7</A>
illustrates the speed of the LAPACK routine for <B><I>LU</I></B> factorization of a
real matrix<A NAME="7805"></A>,
DGETRF<A NAME="7806"></A> in double precision.
This corresponds to 64-bit floating-point arithmetic.
A block size of 1 means that the unblocked algorithm
is used, since it is faster than -- or at least as fast as -- a
blocked algorithm. These numbers may be compared to those for DGEMM in
Table <A HREF="node71.html#emmtable">3.12</A>, which should be upper bounds.
<P>
<BR>
<DIV ALIGN="CENTER">
<A NAME="tablu"></A>
<DIV ALIGN="CENTER">
<A NAME="7809"></A>
<TABLE CELLPADDING=3 BORDER="1">
<CAPTION><STRONG>Table 3.7:</STRONG>
Speed in megaflops of DGETRF for square matrices of order <B><I>n</I></B></CAPTION>
<TR><TD ALIGN="LEFT"> </TD>
<TD ALIGN="CENTER">No. of</TD>
<TD ALIGN="CENTER">Block</TD>
<TD ALIGN="CENTER" COLSPAN=2>Values of <B><I>n</I></B></TD>
</TR>
<TR><TD ALIGN="LEFT"> </TD>
<TD ALIGN="CENTER">processors</TD>
<TD ALIGN="CENTER">size</TD>
<TD ALIGN="RIGHT">100</TD>
<TD ALIGN="RIGHT">1000</TD>
</TR>
<TR><TD ALIGN="LEFT">Dec Alpha Miata</TD>
<TD ALIGN="CENTER">1</TD>
<TD ALIGN="CENTER">28</TD>
<TD ALIGN="RIGHT">172</TD>
<TD ALIGN="RIGHT">370</TD>
</TR>
<TR><TD ALIGN="LEFT">Compaq AlphaServer DS-20</TD>
<TD ALIGN="CENTER">1</TD>
<TD ALIGN="CENTER">28</TD>
<TD ALIGN="RIGHT">353</TD>
<TD ALIGN="RIGHT">440</TD>
</TR>
<TR><TD ALIGN="LEFT">IBM Power 3</TD>
<TD ALIGN="CENTER">1</TD>
<TD ALIGN="CENTER">32</TD>
<TD ALIGN="RIGHT">278</TD>
<TD ALIGN="RIGHT">551</TD>
</TR>
<TR><TD ALIGN="LEFT">IBM PowerPC</TD>
<TD ALIGN="CENTER">1</TD>
<TD ALIGN="CENTER">52</TD>
<TD ALIGN="RIGHT">77</TD>
<TD ALIGN="RIGHT">148</TD>
</TR>
<TR><TD ALIGN="LEFT">Intel Pentium II</TD>
<TD ALIGN="CENTER">1</TD>
<TD ALIGN="CENTER">40</TD>
<TD ALIGN="RIGHT">132</TD>
<TD ALIGN="RIGHT">250</TD>
</TR>
<TR><TD ALIGN="LEFT">Intel Pentium III</TD>
<TD ALIGN="CENTER">1</TD>
<TD ALIGN="CENTER">40</TD>
<TD ALIGN="RIGHT">143</TD>
<TD ALIGN="RIGHT">297</TD>
</TR>
<TR><TD ALIGN="LEFT">SGI Origin 2000</TD>
<TD ALIGN="CENTER">1</TD>
<TD ALIGN="CENTER">64</TD>
<TD ALIGN="RIGHT">228</TD>
<TD ALIGN="RIGHT">452</TD>
</TR>
<TR><TD ALIGN="LEFT">SGI Origin 2000</TD>
<TD ALIGN="CENTER">4</TD>
<TD ALIGN="CENTER">64</TD>
<TD ALIGN="RIGHT">190</TD>
<TD ALIGN="RIGHT">699</TD>
</TR>
<TR><TD ALIGN="LEFT">Sun Ultra 2</TD>
<TD ALIGN="CENTER">1</TD>
<TD ALIGN="CENTER">64</TD>
<TD ALIGN="RIGHT">121</TD>
<TD ALIGN="RIGHT">240</TD>
</TR>
<TR><TD ALIGN="LEFT">Sun Enterprise 450</TD>
<TD ALIGN="CENTER">1</TD>
<TD ALIGN="CENTER">64</TD>
<TD ALIGN="RIGHT">163</TD>
<TD ALIGN="RIGHT">334</TD>
</TR>
</TABLE>
</DIV>
</DIV>
<BR>
<P>
Table <A HREF="node68.html#tabchol">3.8</A>
gives similar results for Cholesky factorization<A NAME="7822"></A>.
<P>
<BR>
<DIV ALIGN="CENTER">
<A NAME="tabchol"></A>
<DIV ALIGN="CENTER">
<A NAME="7824"></A>
<TABLE CELLPADDING=3 BORDER="1">
<CAPTION><STRONG>Table 3.8:</STRONG>
Speed in megaflops of DPOTRF for matrices of order <B><I>n</I></B> with UPLO =
`U'</CAPTION>
<TR><TD ALIGN="LEFT"> </TD>
<TD ALIGN="CENTER">No. of</TD>
<TD ALIGN="CENTER">Block</TD>
<TD ALIGN="CENTER" COLSPAN=2>Values of <B><I>n</I></B></TD>
</TR>
<TR><TD ALIGN="LEFT"> </TD>
<TD ALIGN="CENTER">processors</TD>
<TD ALIGN="CENTER">size</TD>
<TD ALIGN="RIGHT">100</TD>
<TD ALIGN="RIGHT">1000</TD>
</TR>
<TR><TD ALIGN="LEFT">Dec Alpha Miata</TD>
<TD ALIGN="CENTER">1</TD>
<TD ALIGN="CENTER">28</TD>
<TD ALIGN="RIGHT">197</TD>
<TD ALIGN="RIGHT">399</TD>
</TR>
<TR><TD ALIGN="LEFT">Compaq AlphaServer DS-20</TD>
<TD ALIGN="CENTER">1</TD>
<TD ALIGN="CENTER">28</TD>
<TD ALIGN="RIGHT">306</TD>
<TD ALIGN="RIGHT">464</TD>
</TR>
<TR><TD ALIGN="LEFT">IBM Power 3</TD>
<TD ALIGN="CENTER">1</TD>
<TD ALIGN="CENTER">32</TD>
<TD ALIGN="RIGHT">299</TD>
<TD ALIGN="RIGHT">586</TD>
</TR>
<TR><TD ALIGN="LEFT">IBM PowerPC</TD>
<TD ALIGN="CENTER">1</TD>
<TD ALIGN="CENTER">52</TD>
<TD ALIGN="RIGHT">79</TD>
<TD ALIGN="RIGHT">125</TD>
</TR>
<TR><TD ALIGN="LEFT">Intel Pentium II</TD>
<TD ALIGN="CENTER">1</TD>
<TD ALIGN="CENTER">40</TD>
<TD ALIGN="RIGHT">118</TD>
<TD ALIGN="RIGHT">253</TD>
</TR>
<TR><TD ALIGN="LEFT">Intel Pentium III</TD>
<TD ALIGN="CENTER">1</TD>
<TD ALIGN="CENTER">40</TD>
<TD ALIGN="RIGHT">142</TD>
<TD ALIGN="RIGHT">306</TD>
</TR>
<TR><TD ALIGN="LEFT">SGI Origin 2000</TD>
<TD ALIGN="CENTER">1</TD>
<TD ALIGN="CENTER">64</TD>
<TD ALIGN="RIGHT">222</TD>
<TD ALIGN="RIGHT">520</TD>
</TR>
<TR><TD ALIGN="LEFT">SGI Origin 2000</TD>
<TD ALIGN="CENTER">4</TD>
<TD ALIGN="CENTER">64</TD>
<TD ALIGN="RIGHT">137</TD>
<TD ALIGN="RIGHT">1056</TD>
</TR>
<TR><TD ALIGN="LEFT">Sun Ultra 2</TD>
<TD ALIGN="CENTER">1</TD>
<TD ALIGN="CENTER">64</TD>
<TD ALIGN="RIGHT">131</TD>
<TD ALIGN="RIGHT">276</TD>
</TR>
<TR><TD ALIGN="LEFT">Sun Enterprise 450</TD>
<TD ALIGN="CENTER">1</TD>
<TD ALIGN="CENTER">64</TD>
<TD ALIGN="RIGHT">178</TD>
<TD ALIGN="RIGHT">391</TD>
</TR>
</TABLE>
</DIV>
</DIV>
<BR>
<P>
LAPACK, like LINPACK, provides a factorization for symmetric
indefinite<A NAME="7836"></A>
matrices, so that <B><I>A</I></B> is factorized as <B><I>P U D U</I><SUP><I>T</I></SUP> <I>P</I><SUP><I>T</I></SUP></B>, where <B><I>P</I></B> is a
permutation matrix, and <B><I>D</I></B> is block diagonal with blocks of order 1
or 2. A block form of this algorithm has been derived,
and is implemented in the LAPACK routine
SSYTRF<A NAME="7837"></A>/DSYTRF<A NAME="7838"></A>.
It has to duplicate a little of the computation in order
to ``look ahead''
to determine the necessary row and column interchanges, but the extra work
can be more than compensated for by the greater speed of updating the matrix
by blocks as is illustrated in Table <A HREF="node68.html#tabbk"
NAME="7840">3.9</A>,
provided that <B><I>n</I></B> is large enough.
<P>
<BR>
<DIV ALIGN="CENTER">
<A NAME="tabbk"></A>
<DIV ALIGN="CENTER">
<A NAME="7842"></A>
<TABLE CELLPADDING=3 BORDER="1">
<CAPTION><STRONG>Table 3.9:</STRONG>
Speed in megaflops of DSYTRF for matrices of order <B><I>n</I></B> with UPLO =
`U' on an IBM Power 3</CAPTION>
<TR><TD ALIGN="CENTER">Block</TD>
<TD ALIGN="CENTER" COLSPAN=2>Values of <B><I>n</I></B></TD>
</TR>
<TR><TD ALIGN="CENTER">size</TD>
<TD ALIGN="RIGHT">100</TD>
<TD ALIGN="RIGHT">1000</TD>
</TR>
<TR><TD ALIGN="CENTER">1</TD>
<TD ALIGN="RIGHT">186</TD>
<TD ALIGN="RIGHT">215</TD>
</TR>
<TR><TD ALIGN="CENTER">32</TD>
<TD ALIGN="RIGHT">130</TD>
<TD ALIGN="RIGHT">412</TD>
</TR>
</TABLE>
</DIV>
</DIV>
<BR>
<P>
LAPACK, like LINPACK, provides <B><I>LU</I></B> and Cholesky factorizations of
band matrices. The LINPACK algorithms can easily be restructured to use
Level 2 BLAS, though that has little effect on performance for
matrices of very narrow bandwidth. It is also possible to use Level 3 BLAS,
at the price of doing some extra work with zero elements outside the band
[<A
HREF="node151.html#lapwn21">48</A>]. This becomes worthwhile for matrices of large order and
semi-bandwidth greater than 100 or so.
<P>
<P>
<HR>
<!--Navigation Panel-->
<A NAME="tex2html5115"
HREF="node69.html">
<IMG WIDTH="37" HEIGHT="24" ALIGN="BOTTOM" BORDER="0" ALT="next"
SRC="next_motif.png"></A>
<A NAME="tex2html5109"
HREF="node67.html">
<IMG WIDTH="26" HEIGHT="24" ALIGN="BOTTOM" BORDER="0" ALT="up"
SRC="up_motif.png"></A>
<A NAME="tex2html5103"
HREF="node67.html">
<IMG WIDTH="63" HEIGHT="24" ALIGN="BOTTOM" BORDER="0" ALT="previous"
SRC="previous_motif.png"></A>
<A NAME="tex2html5111"
HREF="node1.html">
<IMG WIDTH="65" HEIGHT="24" ALIGN="BOTTOM" BORDER="0" ALT="contents"
SRC="contents_motif.png"></A>
<A NAME="tex2html5113"
HREF="node152.html">
<IMG WIDTH="43" HEIGHT="24" ALIGN="BOTTOM" BORDER="0" ALT="index"
SRC="index_motif.png"></A>
<BR>
<B> Next:</B> <A NAME="tex2html5116"
HREF="node69.html">QR Factorization</A>
<B> Up:</B> <A NAME="tex2html5110"
HREF="node67.html">Examples of Block Algorithms</A>
<B> Previous:</B> <A NAME="tex2html5104"
HREF="node67.html">Examples of Block Algorithms</A>
  <B> <A NAME="tex2html5112"
HREF="node1.html">Contents</A></B>
  <B> <A NAME="tex2html5114"
HREF="node152.html">Index</A></B>
<!--End of Navigation Panel-->
<ADDRESS>
<I>Susan Blackford</I>
<BR><I>1999-10-01</I>
</ADDRESS>
</BODY>
</HTML>
|
