forked from Reference-ScaLAPACK/scalapack-www
-
Notifications
You must be signed in to change notification settings - Fork 0
/
scalapack-2.0.txt
282 lines (214 loc) · 9.31 KB
/
scalapack-2.0.txt
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
ScaLAPACK 2.0
=============
Release date: Fr 11/11/11.
This material is based upon work supported by the National Science Foundation
and the Department of Energy under Grant No. NSF-OCI-1032861, NSF-CCF-00444486,
NSF-CNS 0325873, NSF-EIA 0122599, NSF-ACI-0090127, DOE-DE-FC02-01ER25478,
DOE-DE-FC02-06ER25768.
ScaLAPACK is a software package provided by Univ. of Tennessee, Univ. of
California, Berkeley, and Univ. of Colorado Denver.
1. Support and questions
------------------------
http://icl.cs.utk.edu/lapack-forum/
2. ScaLAPACK 2.0: What's new
----------------------------
1. *PxHSEQR: Nonsymmetric Eigenvalue Problem.*
Contribution by Robert Granat, Bo Kågström, Meiyue Shao (Umeå
University and HPC2N), and Daniel Kressner (École Polytechnique
Fédérale de Lausanne).
Compute the eigenvalues of a nonsymmetric real matrix. Implement the parallel
distributed Hessenberg QR algorithm with the small bulge multi-shift QR
algorithm together with aggressive early deflation.
See reference <<1>> below.
2. *PxSYEVR/PxHEEVR: MRRR (Multiple Relatively Robust Representations) algorithm.*
Contribution by Christof Voemel.
Compute the eigenvalues of a symmetric/Hermitian matrix. Implement the parallel
distributed MRRR algorithm.
See reference <<2>> below.
3. *BLACS revamping.*
Contribution by Rodney James (UC Denver).
3. References
-------------
[1] [[1]]:: Robert Granat, Bo Kågström, and Daniel Kressner.
A Novel Parallel QR Algorithm for Hybrid Distributed Memory HPC Systems.
SIAM J. Sci. Comput. 32(4):2345-2378, 2010.
[2] [[2]]:: Christof Voemel. ScaLAPACK's MRRR Algorithm. ACM Transactions on
Mathematical Software, 37(1):1-35, 2010.
4. External Contributors
------------------------
* Robert Granat, Umeå University and HPC2N
* Bo Kågström, Umeå University and HPC2N
* Daniel Kressner, École Polytechnique Fédérale de Lausanne
* Meiyue Shao, Umeå University and HPC2N
* Christof Voemel
5. Thanks
---------
Thanks for bug-report/patches/suggestions to: Yana Gurieva (Intel MKL), dstrubbe, Jill Reese (MathWorks), François Thomas, Sebastian Tschiatschek and Joost VandeVondele.
6. Developer list
-----------------
.Principal Investigators
* Jim Demmel (University of California, Berkeley, USA)
* Jack Dongarra (University of Tennessee and ORNL, USA)
* Julien Langou (University of Colorado Denver, USA)
.LAPACK developers involved in this release
* Rodney James (University of Colorado Denver, USA)
* Julie Langou (University of Tennessee, USA)
7. More details
----------------
7.1. PxHSEQR: Nonsymmetric Eigenvalue Problem
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
.contributors
Robert Granat, Umeå University and HPC2N,
Bo Kågström, Umeå University and HPC2N,
Daniel Kressner, École Polytechnique Fédérale de Lausanne, and
Meiyue Shao, Umeå University and HPC2N
.lapacker
Rodney James (UC Denver)
.contribution
Compute the eigenvalues of a nonsymmetric real matrix. Implement the parallel
distributed Hessenberg QR algorithm with the small bulge multi-shift QR
algorithm together with aggressive early deflation.
- PxHSEQR computes the eigenvalues of an upper Hessenberg matrix H and,
optionally, the matrices T and Z from the Schur decomposition H = Z*T*Z^T^,
where T is an upper quasi-triangular matrix (the Schur form), and Z is the
orthogonal matrix of Schur vectors. Optionally Z may be postmultiplied into an
input orthogonal matrix Q so that this routine can give the Schur factorization
of a matrix A which has been reduced to the Hessenberg form H by the orthogonal
matrix Q: A = Q*H*Q^T^ = (QZ)\*T*(QZ)^T^.
- PxGEBAL balances a general real matrix A. This involves, first, permuting A by
a similarity transformation to isolate eigenvalues on the diagonal; and second,
applying a diagonal similarity transformation to make the rows and columns as
close in norm as possible. Both steps are optional. Balancing may reduce the
1-norm of the matrix, and improve the accuracy of the computed eigenvalues
and/or eigenvectors.
.reference
Robert Granat, Bo Kågström, and Daniel Kressner. A Novel Parallel
QR Algorithm for Hybrid Distributed Memory HPC Systems. SIAM J. Sci. Comput.
32(4):2345-2378, 2010.
.computational subroutines
pdhseqr.f pshseqr.f
pdgebal.f psgebal.f
[source,fortran]
----
SUBROUTINE PDHSEQR( JOB, COMPZ, N, ILO, IHI, H, DESCH, WR, WI, Z,
$ DESCZ, WORK, LWORK, IWORK, LIWORK, INFO )
*
* .. Scalar Arguments ..
INTEGER IHI, ILO, INFO, LWORK, LIWORK, N
CHARACTER COMPZ, JOB
* ..
* .. Array Arguments ..
INTEGER DESCH( * ) , DESCZ( * ), IWORK( * )
DOUBLE PRECISION H( * ), WI( N ), WORK( * ), WR( N ), Z( * )
----
[source,fortran]
----
SUBROUTINE PDGEBAL( JOB, N, A, DESCA, ILO, IHI, SCALE, INFO )
*
IMPLICIT NONE
*
* .. Scalar Arguments ..
CHARACTER JOB
INTEGER IHI, ILO, INFO, N
* ..
* .. Array Arguments ..
INTEGER DESCA( * )
DOUBLE PRECISION A( * ), SCALE( * )
----
.auxiliary subroutines
bdlaapp.f bslaapp.f
bdlaexc.f bslaexc.f
bdtrexc.f bstrexc.f
dlaqr6.f slaqr6.f
pbdtrord.f pbstrord.f
pbdtrsen.f pbstrsen.f
pdlaqr0.f pslaqr0.f
pdlaqr1.f pslaqr1.f
pdlaqr2.f pslaqr2.f
pdlaqr3.f pslaqr3.f
pdlaqr4.f pslaqr4.f
pdlaqr5.f pslaqr5.f
pdrot.f psrot.f
pdhseqr.f pshseqr.f
pdlamve.f pslamve.f
pdgebal.f psgebal.f
pilaenvx.f
piparmq.f
[[pqr]]
.Performance of previous ScaLAPACK implementation PDLAHQR (red left bars) vs. new implementation PDHSEQR (blue right bars)
image::newpqr.png[]
The graph is extracted from reference <<1>> and the experiments have been performed
by Robert Granat, Bo Kågström, and Daniel Kressner.
This is using Intel Xeon quadcore nodes when computing the Schur form of a dense random matrix reduced to
Hessenberg form. (a) Execution times for a 4000 x 4000 matrix using 1 or 4
cores of a single node. (b) Execution times for a 16 000 x 16 000 matrix using
100 cores of 25 nodes.
7.2. PxSYEVR/PxHEEVR: MRRR (Multiple Relatively Robust Representations) algorithm
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
.contributors
Christof Voemel
.lapacker
Rodney James (UC Denver)
.contribution
This parallel algorithm is derived from Parlett and Dhillon's SIAG-LA
prize-winning work on sequential MRRR. Compared to other algorithms, parallel
MRRR has some striking advantages. First, for an n x n matrix on p processors,
a tridiagonal inverse iteration can require up to O(n^3^) operations and
O(n^2^) memory on a single processor to guarantee the correctness of the
computed eigenpairs. MRRR is guaranteed to produce the right answer with
O(n^2^/p) memory, and it does not need reorthogonalization. Second, MRRR allows
the computation of subsets at reduced cost, whereas QR and Divide & Conquer do
not. For computing k eigenpairs, the tridiagonal parallel MRRR requires O(nk/p)
operations per processor.
.reference
Christof Voemel. ScaLAPACK's MRRR Algorithm. ACM Transactions on Mathematical
Software, 37(1):1-35, 2010.
.driver subroutines
pssyevr.f pdsyevr.f pcheevr.f pzheevr.f
[source,fortran]
----
SUBROUTINE PDSYEVR( JOBZ, RANGE, UPLO, N, A, IA, JA,
$ DESCA, VL, VU, IL, IU, M, NZ, W, Z, IZ,
$ JZ, DESCZ, WORK, LWORK, IWORK, LIWORK,
$ INFO )
*
* .. Scalar Arguments ..
CHARACTER JOBZ, RANGE, UPLO
INTEGER IA, IL, INFO, IU, IZ, JA, JZ, LIWORK, LWORK, M,
$ N, NZ
DOUBLE PRECISION VL, VU
* ..
* .. Array Arguments ..
INTEGER DESCA( * ), DESCZ( * ), IWORK( * )
DOUBLE PRECISION A( * ), W( * ), WORK( * ), Z( * )
----
.auxiliary subroutines
slar1v.f dlar1v.f
slar1va.f dlar1va.f
slarra.f dlarra.f
slarrb.f dlarrb.f
slarrb2.f dlarrb2.f
slarrc.f dlarrc.f
slarrd.f dlarrd.f
slarrd2.f dlarrd2.f
slarre2.f dlarre2.f
slarre2a.f dlarre2a.f
slarrf.f dlarrf.f
slarrf2.f dlarrf2.f
slarrk.f dlarrk.f
slarrv.f dlarrv.f
slarrv2.f dlarrv2.f
sstegr2.f dstegr2.f
sstegr2a.f dstegr2a.f
sstegr2b.f dstegr2b.f
[[syev]]
.Performance of ScaLAPACK implementation PDSYEV (red) and PDSYEVD (green) vs. the new PDSYEVR (blue)
image::syev.png[]
7.3. BLACS revamping
~~~~~~~~~~~~~~~~~~~~~
.lapacker
Rodney James (UC Denver)
.contribution
With ScaLAPACK 2.0, the (MPI) BLACS is now completely integrated into
ScaLAPACK. Linking a ScaLAPACK application now only requires linking with
libscalapack.a, liblapack.a, libblas.a, and possibly the MPI libraries.