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
|
This file is outdated and not very comprehensive. ATLAS installation is
covered in detail in ATLAS/doc/atlas_install.pdf.
Before doing anything, scope the ATLAS errata file for known errors/problems:
http://math-atlas.sourceforge.net/errata.html
and apply any bug fixes/workarounds you find there.
Note that the documentation on the website will repeat most of this
information, and will be much more current. The docs in this tarfile
are here mainly for convenience and for users not connected to the net.
Others should scope the website for the most current documentation:
http://math-atlas.sourceforge.net/faq.html#doc
If you are a Windows user please read ATLAS/doc/Windows.txt before proceeding.
There are two mandatory steps to ATLAS installation (config & build), as
well as three optional steps (test, time, install) and these steps are
described in detail below. For the impatient, here is the basic outline:
**************************************************
mkdir my_build_dir ; cd my_build_dir
/path/to/ATLAS/configure [flags]
make ! tune and compile library
make check ! perform sanity tests
make ptcheck ! checks of threaded code for multiprocessor systems
make time ! provide performance summary as % of clock rate
make install ! Copy library and include files to other directories
**************************************************
If you want to build a dynamic/shared library, see below at header:
BUILDING DYNAMIC/SHARED LIBRARIES
If you want to build a full LAPACK library (i.e. all of the lapack library,
including those lapack routines not natively provided by ATLAS), see:
NOTE ON BUILDING A FULL LAPACK LIBRARY
********** Important Install Information: CPU THROTTLING ***********
Most OSes (including Linux) now turn on CPU throttling for power management
**even if you are using a desktop**. CPU throttling makes pretty much all
timings completely random, and so any ATLAS install will be junk. Therefore,
before installing ATLAS, turn off CPU throttling. For most PCs, you can
switch it off in the BIOS (eg., on my Athlon-64 machine, I can say "No" to
"Cool and Quiet" under "Power Management"). Most OSes also provide a way
to switch off CPU throttling, but that varies from OS to OS. Under Fedora,
at any rate, the following command seemed to work:
/usr/bin/cpufreq-selector -g performance
On my Core2Duo, cpufreq-selector only changes the parameters of the first CPU,
regardless if which cpu you specify. I suspect this is a bug, because on
earlier systems, the remaining CPUs were controlled via a logical link to
/sys/devices/system/cpu/cpu0/. In this case, the only way I found to force
the second processor to also run at its peak frequency was to issue the
following as root after setting CPU0 to performance:
cp /sys/devices/system/cpu/cpu0/cpufreq/scaling_governor \
/sys/devices/system/cpu/cpu1/cpufreq/scaling_governor
More modern and non-broken systems use cpufreq-set, with -c <#> appended
to change the performance of each core in turn. For example, to speedup both
processors of a dual system you would issue:
/usr/bin/cpufreq-set -g performance -c 0
/usr/bin/cpufreq-set -g performance -c 1
On Kubuntu, I had problems with this not working because scaling_max_freq
was set to the minimal speed. To fix, I had to first increase the max scaling
frequency, which you can do (as root) by (where <#> below is replaced by each processor
number eg., 0 and 1 for dual processor system):
cd /sys/devices/system/cpu/cpu<#>/cpufreq
cp cpuinfo_max_freq scaling_max_freq
In Kubuntu 9.10, the only fix I found was to issue:
sudo echo "performance" > \
/sys/devices/system/cpu/cpuX/cpufreq/scaling_governor
Where 'X' is replaced by each of your cpu numbers in turn (eg., if you have
a quad processor, you would issue this command four times, using X=[0,1,2,3]).
Under MacOS or Windows, you may be able to change this under the power settings.
I have reports that issuing "powercfg /q" in cmd.exe under windows will tell
you whether windows is throttling or not.
ATLAS config tries to detect if CPU throttling is enabled, but it may not
always detect it, and sometimes may detect it after you have disabled it.
In the latter case, to force the configure to continue regardless of the
results of the CPU throttling probe, read the error message you get when
ATLAS's configure dies upon detecting throttling.
********************************** CONFIG *************************************
First, create a directory where you will build ATLAS. It can be anywhere in
your filesystem, and does not need to be under the ATLAS/ subdirectory
(though it can be if you like). In general, giving it a descriptive name
is good, for instance:
mkdir ATLAS_Linux_P4E ; cd ATLAS_Linux_P4E
From this directory, configure the new directory to build atlas with the
command:
/path/to/ATLAS/configure [flags]
Obviously, "/path/to/ATLAS" is the full or relative path from where you are
to the directory created by the ATLAS tarfile. To see a list of available
flags to configure type "path/to/ATLAS/configure --help". In general, no
flags are required, but there are many useful flags for helping with various
problems, in particular changing compilers, setting 32 or 64 bit libraries,
etc. To ensure building 32 bit libraries, add the flag -b 32, and -b 64
to force 64 bit libraries (must be using a 64 bit-capable compiler on a
64-bit Operating System).
If the machine you are using is not heavily loaded, then you can vastly
improve the accuracy of ATLAS's timings by using the wall timer rather
than the CPU timer. If you are on a x86, you can get cycle accurate timings
by throwing the flag:
-D c -DPentiumCPS=<your Mhz>
So, for my 2.4Ghz machine, I would specify:
-D c -DPentiumCPS=2400
Under Linux, you can find a decent estimate of your clock rate by
cat /proc/cpuinfo
If you can't use the cycle-accurate wall timer, the normal WALL timer
is still much more accurate than the CPU timer, and you can use it by
adding this to configure:
-D c -DWALL
************ Important Compiler Advice **************
For most systems, ATLAS defaults to using the Gnu compiler collection for
its ATLAS install. This means configure will automatically search for
either g77/gcc or gfortran/gcc. If it can't find them, it will typically
stop with an error message. For some platforms, ATLAS knows good flags
to use for multiple compilers, and so you may get good flags by simply
changing the compiler name. If this doesn't work, you'll need to specify
both the compiler name and the flags to use. For the fortran compiler,
you can switch the fortran compiler without performance or install penalty
on all platforms. To do so, simply add the flags:
-C if <fortran compiler with path> -F if 'fortran flags'
to the configure command. If you need to install ATLAS on a platform that
doesn't have a working fortran compiler, you can do so by adding the flag:
--nof77
instead. Note that the fortran interface to BLAS and LAPACK cannot be built
without a fortran compiler.
Note that all compilers used in an ATLAS install must be able to interoperate.
For more compiler-controlling flags, add --help to the configure command.
*********** Important CLANG Compiler Advice ***********
I have never succesfully built ATLAS with clang. Clang produces the wrong
answer on many kernels, and poorer performance than gcc on every kernel
I've timed. I strongly recommend GNU gcc over Clang/LLVM.
********************************** BUILD **************************************
If config finishes without error, start the build/tuning process by:
make build
(or just "make")
Install times vary widely, depending on whether ATLAS has architectural
defaults for your platform, and the speed of your compiler and computer.
Under gcc/linux, an install may take as little as 15 minutes for all four
types/precisions. On a slow machine lacking architectural defaults, it
However, the user is not required to enter any input during
the build phase, and all operations are logged, so it is safe to start the
install and ignore it until completion.
If you experience problems, read the TROUBLESHOOTING section in
ATLAS/doc/TroubleShoot.txt. ATLAS/README provides an index of all
included ATLAS documentation files.
You should then read ATLAS/doc/TestTime.txt for instructions on testing
and timing your installation.
******************************* SANITY TEST ***********************************
This optional step merely verifies that the built ATLAS libraries are able
to pass basic correctness tests. The standard BLAS testers (i.e. those
that go with the API, as opposed to those written by the ATLAS group) are
written in Fortran77, and so you will need a Fortran compiler installed to
run them. If you have no Fortran77, you can amend the directions below by
prepending "C_" to all target names (eg., "make C_test") to run only those
testers that require a C compiler, but note that in so doing you will get
a less rigorously tested library.
you can run all the sequential sanity tests by:
make check
If you have elected to build the threaded library, you can run the same tests
with the threaded library with:
make ptcheck
A successful sanity test will dump a lot of compilation to the window, followed
with something like:
===========================================================================
DONE BUILDING TESTERS, RUNNING:
SCOPING FOR FAILURES IN BIN TESTS:
fgrep -e fault -e FAULT -e error -e ERROR -e fail -e FAIL \
bin/Linux_PIIISSE1/sanity.out
8 cases: 8 passed, 0 skipped, 0 failed
4 cases: 4 passed, 0 skipped, 0 failed
8 cases: 8 passed, 0 skipped, 0 failed
4 cases: 4 passed, 0 skipped, 0 failed
8 cases: 8 passed, 0 skipped, 0 failed
4 cases: 4 passed, 0 skipped, 0 failed
8 cases: 8 passed, 0 skipped, 0 failed
4 cases: 4 passed, 0 skipped, 0 failed
DONE
SCOPING FOR FAILURES IN CBLAS TESTS:
fgrep -e fault -e FAULT -e error -e ERROR -e fail -e FAIL \
interfaces/blas/C/testing/Linux_PIIISSE1/sanity.out | \
fgrep -v PASSED
make[1]: [sanity_test] Error 1 (ignored)
DONE
SCOPING FOR FAILURES IN F77BLAS TESTS:
fgrep -e fault -e FAULT -e error -e ERROR -e fail -e FAIL \
interfaces/blas/F77/testing/Linux_PIIISSE1/sanity.out | \
fgrep -v PASSED
make[1]: [sanity_test] Error 1 (ignored)
DONE
make[1]: Leaving directory `/home/rwhaley/TEST/TEST/ATLAS3.3.15'
===========================================================================
Note that the "Error 1 (ignored)" is coming from grepping for failure, and
grep is saying it doesn't find any . . .
Assuming you have a fortran compiler, you can also run the full ATLAS
testing scripts, which may take over a day to run. If you have
installed a full LAPACK library, you can also run the standard LAPACK
testers. Please see "EXTENDED ATLAS TESTING" below for more information.
******************* INSTALLING ATLAS FOR MULTIPLE ARCHITECTURES ***************
You can install ATLAS from the same source tree on multiple machines at once
by simply creating different build directories for each architecture.
********************** BUILDING DYNAMIC/SHARED LIBRARIES **********************
ATLAS natively builds to a static library (i.e. libs that usually end in
".a" under unix and ".lib" under windows). ATLAS always builds such a library,
but it can also optionally be requested to build a dynamic/shared library
(typically ending in .so for unix or .dll windows and .dylib for OS X).
To do this for GNU compilers, you typically just need to add
--shared
flag to your configure line.
If you use non-gnu compilers, you'll need to use -Fa to
pass the correct flag(s) to append to force position independent code for
each compiler (don't forget the gcc compiler used in the index files).
NOTE: Since gcc uses one less int register when compiling with this flag, this
could potentially impact performance of the architectural defaults,
but we have not seen it so far.
****************** NOTE ON BUILDING A FULL LAPACK LIBRARY *********************
To build lapack, first download the lapack tarfile from
www.netlib.org/lapack
Then, pass the flag to configure:
--with-netlib-lapack-tarfile=/path/to/downloaded/tarfile
***************************** EXTENDED ATLAS TESTING **************************
ATLAS has two extended testers beyond the sanity checks that can be
automatically invoked from the BLDdir. These tests are longer running and
more complex to interpret than the sanity tests, and so not every user will
want to run them. They are particularly recommended for installers who wish
to use a developer release for production code.
--------------------------------- full_test -----------------------------------
The first is a set of testing scripts written by Antoine Petitet, that
randomly generate testcases for a host of ATLAS's testers. This testing
phase may take as long as two days to complete (and almost always takes
at least 4 hours). To perform this long-running test, simply issue:
make full_test
If you are logged into the host machine remotely, chances are good your
connection will go down before the install completes. Therefore, you
should consider running the above command with nohup.
At the completion of the tests, the extensive output files will be searched
for errors (much as with the sanity tests), and the output sent to the screen.
If you have lost this screen of data, you can regenerate it with the command:
make scope_full_test
Running these tests will create a directory BLDdir/bin/AtlasTest where the
tester resides, and your output files will be stored a $(ARCH) subdir.
If you want to rerun the testers from scratch (rather than just searching
old output), you can simply delete the entire BLDdir/bin/AtlasTest
directory tree, and do "make full_test" again.
----------------------------- lapack_test -------------------------------------
If you have installed the full LAPACK library, then you can run the standard
lapack testers as well. The command you give is:
make lapack_test_[a,s,f]l_[ab,sb,fb,pt]
The first choice (choose one of three) controls which LAPACK library macro is
used in the link for testing:
_l LINK FOR LAPACK Make.inc MACRO
== =================== ==============
a ATLAS's LAPACK $(LAPACKlib)
s system LAPACK $(SLAPACKlib)
f F77 reference LAPACK $(FLAPACKlib)
The second choice (choose one of three) controls which BLAS macros are
used in the link for testing:
_b/pt LINK FOR BLAS Make.inc MACRO
==== ===================== =========================================
ab ATLAS BLAS $(F77BLASlib) $(CBLASlib) $(ATLASlib)
sb system BLAS $(BLASlib)
fb F77 reference BLAS $(FBLASlib)
pt ATLAS' threaded BLAS $(PTF77BLASlib) $(PTCBLASlib) $(ATLASlib)
Not all of these combinations will work without user modification of Make.inc.
You will need to fill in values for
$(BLASlib)
$(SLAPACKlib)
$(FLAPACKlib)
if you want to run the lapack tester against these libraries.
Usually, you will want to test your newly install ATLAS LAPACK & BLAS:
make lapack_test_al_ab
As before, once the testing is complete, you will get the output of a search
for errors though all output files, and you can search them again with:
make scope_lapack_test_al_ab
Unfortunately, the lapack testers always show errors on almost all platforms.
So, how do you know if you have a real error? Real errors will usually
have residuals in the 10^6 range, rather than O(1) (smaller residuals mean
less error). If you are unsure, the best way is to contrast ATLAS with an
all-F77 install:
make lapack_test_fl_fb
(To run this test, you will have to build a stock netlib LAPACK library,
and fill out Make.inc's FLAPACKlib macro appropriately.) You can then see
how the errors reported by ATLAS stack up against the all-F77 version:
if they are roughly the same, then you are usually OK.
All the lapack testers create a directory BLDdir/bin/LAPACK_TEST. For
each test you run there will be a subdirectory
LAOUT_[A,S,F]L_[AB,SB,FB,PT]
where all your output files will be located. Additionally, the results
of the scope (search for error) will be stored in
BLDdir/bin/LAPACK_TEST/SUMMARY_<lapack>_<blas>
Therefore, a typical round of testing might be:
make lapack_test_al_ab
make lapack_test_fl_fb
# compare SUMMARY_al_ab with SUMMARY_fl_fb to check for error
make lapack_test_al_pt
# compare SUMMARY_al_pt with SUMMARY_fl_fb to check for error in parallel lib
If you had an error, you might want to be sure the error was in ATLAS's BLAS
and not lapack, so you could do "make lapack_test_fl_ab", and see if the
error went away. If you filled in the GotoBLAS for the SLAPACKlib & BLASlib
macros, you could scope the error properties of Goto's BLAS and LAPACK.
Many system/vendor LAPACK/BLAS do not provide all of the routines required
to run the LAPACK testers, and some ATLAS testers call ATLAS internal
routines. Therefore, the safest thing if you have missing symbol errors
when building system/vendor tests, is to use ATLAS to pick up any missing
symbols. For instance, here is an example Make.inc output that makes all of
ATLAS testers work with the GotoBLAS on my Athlon-64 workstation:
BLASlib = /opt/lib/libgoto_opteronp-r1.26.a \
$(F77BLASlib) $(CBLASlib) $(ATLASlib)
SLAPACKlib = /opt/lib/libgoto_opteronp-r1.26.a $(FLAPACKlib)
|