Atlas binaries require SSE3 extensions

[Nathaniel Smith]

With stock ATLAS, Python, numpy from natty on x86-32, I get the following behavior when trying to solve this least-squares problem: http://math.nist.gov/MatrixMarket/data/Harwell-Boeing/lsq/well1850.html

$ python -c 'import cPickle; import gzip; import numpy as np; x, y = cPickle.load(gzip.open("well1850.pickle.gz")); np.linalg.lstsq(x, y)'
Illegal instruction

(Note that while this matrix is sparse, I have it stored in a dense numpy array. The file 'well1850.pickle.gz' is attached to make it easier to reproduce.)

Expected behavior: I wouldn't think you should ever get a SIGILL from BLAS, no matter what you do.

I can't seem to find any dbgsym packages for ATLAS, but gdb says:

(gdb) r
Starting program: /usr/bin/python -c import\ cPickle\;\ import\ gzip\;\ import\ numpy\ as\ np\;\ x,\ y\ =\ cPickle.load\(gzip.open\(\"well1850.pickle.gz\"\)\)\;\ np.linalg.lstsq\(x,\ y\)
[Thread debugging using libthread_db enabled]

Program received signal SIGILL, Illegal instruction.
0x00704e89 in ATL_dupMBmm0_4_0_b0 () from /usr/lib/libblas.so.3gf
(gdb) bt
#0 0x00704e89 in ATL_dupMBmm0_4_0_b0 () from /usr/lib/libblas.so.3gf
#1 0x000002c8 in ?? ()
#2 0x0087b4f8 in ?? () from /usr/lib/libblas.so.3gf
Backtrace stopped: previous frame inner to this frame (corrupt stack?)
(gdb)

ProblemType: Bug
DistroRelease: Ubuntu 11.04
Package: libatlas3gf-base 3.8.3-29
ProcVersionSignature: Ubuntu 2.6.38-8.42-generic 2.6.38.2
Uname: Linux 2.6.38-8-generic i686
Architecture: i386
Date: Thu May 5 17:05:48 2011
InstallationMedia: Ubuntu 9.10 "Karmic Koala" - Release i386 (20091028.5)
ProcEnviron:
 LANGUAGE=en_US:en
 PATH=(custom, user)
 LANG=en_US.UTF-8
 SHELL=/usr/bin/zsh
SourcePackage: atlas
UpgradeStatus: Upgraded to natty on 2011-05-05 (0 days ago)

[Pauli Virtanen (pauli-virtanen)]

Worksforme with libatlas3gf-base, but I have a processor with SSE3.

I'm seeing this with libatlas3gf-base: (based on opcodes in the assembler dump of the binaries)

$ detect_cpu_extensions /usr/lib/atlas-base/*.so
/usr/lib/atlas-base/libatlas.so: 3dnow mmx pentium_pro sse sse2 sse3
/usr/lib/atlas-base/libcblas.so: pentium_pro sse sse2
/usr/lib/atlas-base/libf77blas.so: pentium_pro sse sse2
/usr/lib/atlas-base/liblapack_atlas.so: pentium_pro sse sse2

IIRC, Atlas does not have runtime CPU extension detection, so this apparently means the binary shipped with Ubuntu will only work on modern processors with SSE3. I note that currently it seems that te libatlas3gf-base is the only ATLAS package --- the *-sse* versions seem to have been removed.

[Sylvestre Ledru (sylvestre)]

About the drop of optimized packages, see:
http://packages.debian.org/changelogs/pool/main/a/atlas/current/changelog#versionversion3.8.3-25

It is likely to be the same bug as:
http://bugs.debian.org/cgi-bin/bugreport.cgi?bug=622082

[Pauli Virtanen (pauli-virtanen)]

It's an ATLAS issue, so setting the Numpy part to Invalid.

[Nathaniel Smith (njs)]

Yes, indeed, I'm working on a Thinkpad X60s which has a Core Duo -- so, no SSE3. /proc/cpuinfo says:

flags : fpu vme de pse tsc msr pae mce cx8 apic mtrr pge mca cmov clflush dts acpi mmx fxsr sse sse2 ss ht tm pbe nx constant_tsc arch_perfmon bts aperfmperf pni monitor vmx est tm2 xtpr pdcm dts

I understand the rationale for only shipping one Atlas binary, but presumably that one binary should target the lowest common denominator?

[Sylvestre Ledru (sylvestre)]

Yes, I am not arguing. it is a bug.