Package Details: atlas-lapack 1:3.10.3-9

Git Clone URL: https://aur.archlinux.org/atlas-lapack.git (read-only, click to copy)
Package Base: atlas-lapack
Description: Automatically Tuned Linear Algebra Software
Upstream URL: http://math-atlas.sourceforge.net
Licenses: BSD, custom:lapack
Provides: atlas-lapack-base, blas, cblas, lapack
Submitter: ilpianista
Maintainer: henkm
Last Packager: henkm
Votes: 93
Popularity: 0.000050
First Submitted: 2008-04-24 01:36 (UTC)
Last Updated: 2024-02-01 12:28 (UTC)

Required by (550)

Sources (4)

Pinned Comments

phcerdan commented on 2017-06-08 06:48 (UTC) (edited on 2017-06-08 06:49 (UTC) by phcerdan)

Hey I just installed, and make these notes, that might be useful for somebody else: Good explanation in atlas site: http://math-atlas.sourceforge.net/atlas_install/node5.html Follow this, the governor set by cpupower knows shit about CPU without this: http://unix.stackexchange.com/questions/121410/setting-cpu-governor-to-on-demand-or-conservative Summary: http://vincent.jousse.org/tech/archlinux-compile-lapack-atlas-kaldi/ ===========Steps=========== Permanent disable intel_pstate: Edit: /etc/default/grub GRUB_CMDLINE_LINUX_DEFAULT="intel_pstate=disable" and update grub: grub-mkconfig -o /boot/grub/grub.cfg And then enable acpi-cpufreq module: su root echo "acpi-cpufreq" > /etc/modules-load.d/acpi-cpufreq.conf restart. Now cpupower can set frequencies properly. To disable throtling sudo pacman -S cpupower sudo cpupower frequency-set -g performance It should apply to all cores, but if it only apply to the first one: copy files to the other (4 in laptop) sudo cp /sys/devices/system/cpu/cpu0/cpufreq/scaling_governor /sys/devices/system/cpu/cpu1/cpufreq/scaling_governor To restore: sudo cpupower frequency-set -g ondemand If not all cores are set: sudo cp /sys/devices/system/cpu/cpu0/cpufreq/scaling_governor /sys/devices/system/cpu/cpu1/cpufreq/scaling_governor This stuff is only required at build time.

Latest Comments

« First ‹ Previous 1 .. 9 10 11 12 13 14 15 16 17 18 Next › Last »

gdiscry commented on 2011-12-10 14:11 (UTC)

You also forgot to apply the patch here's patch for the PKGBUILD --- atlas-lapack.orig/PKGBUILD 2011-12-10 13:30:42.000000000 +0100 +++ atlas-lapack/PKGBUILD 2011-12-10 15:06:36.050601953 +0100 @@ -15,7 +15,7 @@ provides=("blas" "lapack=$_lapackver" 'cblas') license=('custom:blas' 'custom:lapack' 'custom:atlas') source=(http://www.netlib.org/lapack/lapack-$_lapackver.tgz http://downloads.sourceforge.net/math-atlas/atlas${pkgver}.tar.bz2 blas-license.txt atlas-license.txt lapack.patch makefile.shared.mt makefile.shared.st misdetects_i1_as_i2.patch) -md5sums=('80bf4f7ffacaec71adbb69f6f30af4a3' +md5sums=('02d5706ec03ba885fc246e5fa10d8c70' '1bb3abde499b492b4be1f1a0759fbfa2' '38b6acb8ed5691d25863319d30a8b365' '4903eb06072dfbf94710691ccb6660bf' @@ -64,7 +64,9 @@ LAPACKLIB="$srcdir/lapack-$_lapackver/lapack_LINUX.a" msg 'Build ATLAS' - cd "$srcdir/ATLAS/build" + cd "$srcdir/ATLAS" + patch -uNp0 -i "$srcdir/misdetects_i1_as_i2.patch" + cd "build" rm -rf * ../configure --prefix=/usr/ $ARCHITECTURE_BUILD_OPTS -Fa alg -fPIC \ --with-netlib-lapack="$LAPACKLIB"

gdiscry commented on 2011-12-10 13:56 (UTC)

The MD5 for the LAPACK sources fails. Also, could you create the build directory with mkdir -p so that it doesn't fail if we stop and restart makepkg ?

gborzi commented on 2011-12-10 12:45 (UTC)

Updated with lapack 3.4.0 and the patch suggested by Murmex. Thanks Murmex.

gdiscry commented on 2011-12-10 11:39 (UTC)

Problem solved for the compilation on Intel Core i1 processors, this was actually in the errata: http://math-atlas.sourceforge.net/errata.html#i2i1 Add the following patch: --- ATLAS/CONFIG/src/backend/archinfo_x86.c.orig 2011-12-10 12:03:57.816977701 +0100 +++ ATLAS/CONFIG/src/backend/archinfo_x86.c 2011-12-10 12:04:04.347061389 +0100 @@ -309,9 +309,9 @@ case 0x1A: case 0x1E: case 0x1F: + case 0x25: iret = IntCorei1; break; - case 0x25: case 0x2A: iret = IntCorei2; break;

gdiscry commented on 2011-12-10 09:53 (UTC)

I'm not sure, but looks like Netlib LAPACK 3.4.0 has been released. I have a problem similar to deb's on a Core i5. I found the following bug: http://sourceforge.net/tracker/?func=detail&aid=3384178&group_id=23725&atid=379483 3.8.4 is the latest stable release so I'll try with an unstable release and see if it works. Anyway, providing more information in that bug report might be useful.

gborzi commented on 2011-12-03 11:24 (UTC)

Why was this package flagged out of date?

<deleted-account> commented on 2011-10-25 07:47 (UTC)

Hello, I tried to compile this package on a Core i3-380UM, it failed because of the program "xL1" who was denied some illegal instructions (wonder what it is) maybe this has something to do with compilation options ? I commented out USE_ARCH_DEFAULTS, no effects. I googled this kind of error but did not found any solutions, any hints ? Thanks for this package

gborzi commented on 2011-10-21 18:21 (UTC)

Hi orbisvicis, thanks for the updated PKGBUILD. I'll update a new package, based on your PKGBUILD, but without changing the pkgver number. Compiling it anew is a real pain. Note that the lapack patch is used to remove some object files from the library that contain subroutines already available in atlas. I agree that the performance isn't something to write home about. Try gotoblas2 https://aur.archlinux.org/packages.php?ID=43880 instead. It's much faster than ATLAS on Intel processors and compiles in a much shorter time. On AMD processor gotoblas2 has about the same performance as ATLAS, but still it's easier to compile.

orbisvicis commented on 2011-10-21 17:54 (UTC)

This package is quite a pain to build! I made a few very slight changes to the PKGBUILD @ http://hpaste.org/raw/52921 For example, modern bashisms (which is OK, since makepkg depends explicitly on bash), combined {echo,sed,grep} to sed only, fixed one regex, made the patch name clearer. I think USE_ARCH_DEFAULTS no longer has any effect. It made no difference (performance/time) compiling with or without it, and the docs or the ATLAS source make no mention of it. The installation guide says to use "-Si archdef 0" to ignore architectural defaults. Also, it seems the lapack patch is no longer needed. Builds fine without it... so I commented it out. However, I'm not too happy with the performance (CoreDuo32SSE3) In case anyone is looking for a standard benchmark, HPL (A Portable Implementation of the High-Performance Linpack Benchmark for Distributed-Memory Computers) works very well. I built HPL against netlib-fortran-blas, atlas-fortran-blas, atlas-cblas-nothreads, atlas-cblas-threads and used http://www.advancedclustering.com/faq/how-do-i-tune-my-hpldat-file.html to generate the HPL.dat file, geared towards a single dual-core machine relevant: settings reproduced here 2 # of problems sizes (N) 1000 2000 Ns 1 # of NBs 128 NBs 0 PMAP process mapping (0=Row-,1=Column-major) 1 # of process grids (P x Q) 1 Ps 1 Qs These are my results: netlib-fortran-blas: . 0.97 . 7.93 atlas-cblas-singlethreaded . 0.61 . 4.72 atlas-fortran-blas-multithreaded . 0.41 . 3.02 atlas-cblas-multithreaded . 0.40 . 2.84 The values are pretty consistent across various runs.

awim commented on 2011-07-23 21:36 (UTC)

@gborzi thanks

« First ‹ Previous 1 .. 9 10 11 12 13 14 15 16 17 18 Next › Last »