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
|
2023-03-14 Philippe Michel <philippe.michel7@free.fr>
* Makefile.am, SFMT.c(+), SFMT.h(+), SFMT-common.h(+),
SFMT-params.h(+), SFMT-params19937.h(+), mt19937ar.c(-),
mt19937ar.h(-): replace the original Mersenne Twister RNG by the
improved SFMT version.
2021-10-06 Philippe Michel <philippe.michel7@free.fr>
* neuralnetsse.c, mm_malloc.h(-) Makefile.am, ../configura.ac:
Modify preferred choices for aligned memory allocation. Try
posix_memalign(), then _aligned_malloc(), then a system/compiler
_mm_malloc() ; don't provide the latter any more.
2018-05-12 Philippe Michel <philippe.michel7@sfr.fr>
* simd.h, neuralnetsse.c, neuralnet.c, inputs.c: Add ARM NEON
instructions support.
2016-01-03 Philippe Michel <philippe.michel7@sfr.fr>
* neuralnetsse.c: Use FMA3 instructions in a few places if
available and enabled. The speed gain is modest.
2014-10-12 Philippe Michel <philippe.michel7@sfr.fr>
* neuralnetsse.c: Use knowledge of the limited range of possible
values for base inputs to accelerate neural networks computations.
2014-10-05 Philippe Michel <philippe.michel7@sfr.fr>
* cache.c: Use MurmurHash3 for evaluations cache addressing
instead of the current hashing function. Key computation is faster
and hit rate (very slightly) better.
2013-06-22 Michael Petch <mpetch@gnubg.org>
* ChangeLog, Makefile.am, inputs.c, neuralnet.c, neuralnet.h,
neuralnetsse.c, +simd.h, -sse.h. Rename sse.h to simd.h. Rename
USE_SSE_VECTORIZE to USE_SIMD_INSTRUCTIONS, rename SSE_Supported
to SIMD_Supported, rename DISABLE_SSE_TEST to DISABLE_SIMD_TEST,
rename SSE_CFLAGS to SIMD_CFLAGS, rename file sse.h to simd.h
Make sse/simd usage more consistent
2013-06-15 Philippe Michel <philippe.michel7@sfr.fr>
* neuralnetsse.c, sse.h, inputs.c, ../configure.ac: Use AVX
instructions when built with "configure --enable-simd=avx"
2013-04-15 Philippe Michel <philippe.michel7@sfr.fr>
* cache.c, cache.h, ../eval.c: Use 32 bits variables for keys
instead of long integers. This is what the hashing functions,
taken from an external source, expect.
2011-12-31 Philippe Michel <philippe.michel7@sfr.fr>
* neuralnetsse.c: Fixes error in sigmoid approximation. In
addition, uses reciprocal instead of division only if compiled
with -ffast-math to keep output identical to non-SSE code.
2011-07-14 Philippe Michel <philippe.michel7@sfr.fr>
* cache.c, cache.h: On Intel CPUs, use Gcc builtins for locking
instead of Glib atomic ops. The latter don't work (on RHEL 6) with
more than 20-25 cores. Gcc builtins seem fine up to 48 at least
and bring a small speed improvement in all cases.
2011-05-13 Philippe Michel <philippe.michel7@sfr.fr>
* inputs.c (new file), Makefile.am: baseInputs() computation using
SSE.
2011-04-19 Philippe Michel <philippe.michel7@sfr.fr>
* neuralnet.c, neuralnetsse.c: Reorder frequent tests to put the
most likely branch first.
2011-03-22 Philippe Michel <philippe.michel7@sfr.fr>
* neuralnet.c, neuralnet.h : Remove unused NeuralNetResize(),
simplify NeuralNetCreate(). Use of untrained nets has been removed
for some time.
2011-02-08 Philippe Michel <philippe.michel7@sfr.fr>
* neuralnetsse.c, neuralnet.h : ystein Johansen's patch to handle
a different number of hidden nodes with SSE.
2010-10-29 Philippe Michel <philippe.michel7@sfr.fr>
* neuralnet.c: On FreeBSD, check sse from sysctl
Picked from FreeBSD port patches.
Tue Feb 20 21:25:01 GMT 2007 Jim Segrave <jes@j-e-s.net>
* neuralnet.c: skip sse check if __x86_64 is defined
Fri Apr 29 21:16:41 CEST 2004 Jon Kinsey
* neuralnetsse.c: Split out SSE code to separate file
Wed May 4 18:10:43 GMT 2005 Jim Segrave <jes@jes-2.demon.nl>
* sse.h: add CVS tag
Fri Apr 29 21:16:41 CEST 2004 Jon Kinsey
* neuralnet.c: Minor improvements to new code
* sse.h: New file with sse specific defines
Thu Apr 28 22:35:53 2005 ystein Johansen <oysteim@gnubg.org>
* neuralnet.c (Evaluate128, EvaluateFromBase128, NeuralNetEvaluate128):
New functions for vectorized evaluation of neural networks
with 128 hidden nodes.
Fri Feb 25 11:16:41 CEST 2004 Jon Kinsey
* neuralnet.c Relegate BLAS code as proved slower
Mon May 3 15:16:41 CEST 2004 Joern Thyssen <jth@gnubg.org>
* mt19937ar.c (init_genrand): add mti as parameter to init_genrand.
Thu Apr 22 21:14:32 CEST 2004 Joern Thyssen <jth@gnubg.org>
* mt19937int.[ch]: replaced by mt19937ar.[ch]
* mt19937ar.[ch]: upgrade Mersenne-Twister from
<URL:http://www.math.keio.ac.jp/matumoto/emt.html>.
Thu Dec 18 23:35:48 2003 Joern Thyssen <jth@gnubg.org>
* neuralnet.c: implement Olivier Baur's faster sigmoid
(note that it should probably not be used until the net
has been retrained)
* neuralnet.c (Evaluate, EvaluateFromBase): implement
net propagation as matrix x vector using BLAS from ATLAS.
Fri Jan 10 19:13:15 2003 Joern Thyssen <jth@gnubg.org>
* hash.h: enlage cache to hold cubeful equities as well
(and possibly cubeless ones).
Thu Dec 19 22:53:52 2002 Joern Thyssen <jth@freeshell.org>
* hash.h: add new fields (statistics) to hash.
Mon Aug 5 13:48:34 2002 Gary Wong <gtw@gnu.org>
* mt19937int.c (sgenrand): Fudge the seed so that it fits the
example in the manual more closely.
Mon Jul 29 13:15:22 2002 Gary Wong <gtw@gnu.org>
* mt19937int.c (sgenrand): Upgrade to Makoto Matsumoto's
latest initialisation code. Reported by Eli
Liang <elientzeliang@hotmail.com>.
Thu Jul 4 18:50:27 2002 Jrn Thyssen <jthyssen@dk.ibm.com>
* hash.c (CacheStat): only assign parameters if not NULL.
Wed Jul 3 13:56:25 2002 Joseph Heled <pepster@users.sourceforge.net>
* New cache code. Faster, and higher hit %. You can revert by
compiling with '-DGARY_CACHE'
Sat May 4 13:33:14 2002 Joseph Heled <pepster@users.sourceforge.net>
* neuralnet.c (NeuralNetCreateDirect): add initialization for
savedBase, savedIBase.
Fri Apr 26 19:23:53 2002 ystein Johansen <oeysteij@online.no>
* neuralnet.c: Add neural net speedup from fibs2html
Wed Feb 6 10:03:53 2002 Gary Wong <gtw@research.att.com>
* neuralnet.c (CheckRC): Use time() as a seed for irandinit().
Fri Feb 1 10:49:00 2002 Gary Wong <gtw@research.att.com>
* neuralnet.c (NeuralNetCreateDirect): New function.
Thu Aug 23 10:57:16 2001 Gary Wong <gtw@research.att.com>
* dynarray.c, hash.c, heap.c, list.c: #include <stddef.h>, to get
definition of NULL.
Thu Apr 5 14:20:57 2001 Gary Wong <gtw@research.att.com>
* dynarray.c (DynArrayAdd): Fix realloc size parameter.
Mon Mar 19 10:47:31 2001 Gary Wong <gtw@research.att.com>
* hash.c: Ignore NULL parameters.
Thu Feb 22 11:42:57 2001 Gary Wong <gtw@research.att.com>
* buffer.c: Check whether <sys/uio.h> exists.
Tue Feb 13 13:12:19 2001 Gary Wong <gtw@research.att.com>
* neuralnet.c (NeuralNetCreate): Fix initial random weights
(allow negative values).
Tue Jan 16 13:34:39 2001 Gary Wong <gtw@research.att.com>
* hash.c (CacheResize): Recreate the original size if the new
allocation fails.
Tue Nov 14 13:03:23 2000 Gary Wong <gtw@research.att.com>
* isaacs.h: Remove definitions which polluted the namespace
(and weren't even used).
Fri Nov 10 13:45:53 2000 Gary Wong <gtw@research.att.com>
* neuralnet.c (sigmoid): Optimisations.
(NeuralNetCreate, NeuralNetResize): Use irand() instead of rand_r(),
for better quality random numbers and improved portability.
Fri Sep 29 14:08:05 2000 Joseph Heled <pepe@internet.co.nz>
* config.h: undef VERSION and PACKAGE before setting them.
Wed Aug 2 13:26:08 2000 Joseph Heled <pepe@internet.co.nz>
* neuralnet.c: Improve sigmoid accuracy by using a table lookup for
the first two digits and using (exp(x) = 1 + x) for x < 0.1.
Thu Jul 27 18:20:58 2000 Gary Wong <gary@cs.arizona.edu>
* event.c: Define nothing on systems that can't support it.
* buffer.c: Define nothing on systems that can't support it.
* fifo.c (FifoIOVec): Do not define on systems that can't support it.
Sat Mar 11 20:09:10 2000 Gary Wong <gary@cs.arizona.edu>
* neuralnet.c (NeuralNetDifferentiate): New function.
Fri Jan 14 09:34:52 2000 Gary Wong <gary@cs.arizona.edu>
* fifo.c: replaced bcopy with memcpy.
Sat Jan 8 17:57:54 2000 Gary Wong <gary@cs.arizona.edu>
* event.c: removed syslog code (unportable and unnecessary).
Thu Jan 6 22:24:39 2000 Gary Wong <gary@cs.arizona.edu>
* neuralnet.c (NeuralNetLoadBinary, NeuralNetLoadBinary): New
functions.
Mon Jan 3 03:40:48 2000 Gary Wong <gary@cs.arizona.edu>
* rand_r.c (rand_r): New function.
Sun Jan 2 11:33:58 2000 Gary Wong <gary@cs.arizona.edu>
* hash.c (CacheResize): New function.
|
