On Sun, 10 Apr 2005, M. Edward (Ed) Borasky wrote:
Peter Dalgaard wrote:
I think it's an Atlas issue, and possibly an Atlas/Athlon32 issue. The built-in BLAS in R-beta don't show this. Is there enough detail on what's happening available for me to take this to the Atlas folks? They've done a lot of work, including assembler code, on Athlons and their 64-bit descendents. My "main system", where I've been doing this testing, is a rather old Athlon T-Bird."M. Edward (Ed) Borasky" <[EMAIL PROTECTED]> writes:
[EMAIL PROTECTED] wrote:
This looks more serious. 100 times machine precision is quite a large margin in these matters. Could you perhaps stick in a printout of the two terms and their difference?
I have an ATLAS build on AMD64 and it passes all the checks, but it is using ATLAS 3.7.8, so you might want to try an upgrade.
Attached ... you actually weren't very far off:
...
1 2 3 4 7 8 9 32.971099 37.113091 27.472204 16.891921 32.320560 -6.091053 -26.953745 12 13 14 15 16 17 18print (f2[common])
...
41.798651 40.734935 40.285066 24.876177 8.442082 46.373463 72.652242 118 120 121 122 123 124 125 65.983901 81.140660 101.389698 92.784665 86.803528 66.813059 76.464152 126 127 128 129 130 131 132 85.562396 80.164720 55.046451 22.602751 38.602215 35.466808 28.565003 133 134 135 136 137 138 139 30.487396 27.515347 17.475536 49.119123 11.994736 14.701687 49.795201 140 141 142 143 144 145 146 5.664599 24.711067 20.426534 53.013693 5.758723 19.324367 41.190110 147 148 149 151 152 153 14.189862 -19.275130 35.155615 20.525269 40.584670 18.702940...
Aha! 100 times machine precision in not all that much when the numbers themselves are in double digits. In fact, one is over 100. The case that triggers the failure is #149
147 148 149 151 152 -1.598721e-14 -1.065814e-14 -2.842171e-14 -1.065814e-14 -2.131628e-14
which is 2 ULP off by my reckoning (scaling 35.15 to be between 0.5 and 1 makes the error 2.842e-14/64 = 4.44e-16 and [EMAIL PROTECTED] is 2.22e-16).
So again, we might be too strict. I just wonder why we haven't heard of this on any other platforms.
Incidentally, even though Atlas does have most parameters set to pre-defined values for the Athlon/Linux, it does in fact make *some* decisions when it compiles, which may be why I'm showing this and nobody else is. I'm using Gentoo Linux, which recompiles nearly everything from source, including Atlas and R, when it does an install. Most of the other Linux distros have pre-compiled binaries for the various packages. Debian, for example, has pre-compiled Atlas libraries for P3, P4 and Athlon.
I think the issue is ATLAS on your old Athlon. ATLAS 3.6.0 compiled from the sources works correctly with gcc-3.4.3 on my Athlon MP (and also on an Athlon XP), but AFAIR those have instructions the Athlon Thunderbird does not have. (Both my machines with such Athlons fried their motherboards, so I no longer have access to one.)
Incidentally to Peter: ATLAS 3.7.8 is an unreleased unstable version, so I would hesitate to recommend it over 3.6.0.
-- Brian D. Ripley, [EMAIL PROTECTED] Professor of Applied Statistics, http://www.stats.ox.ac.uk/~ripley/ University of Oxford, Tel: +44 1865 272861 (self) 1 South Parks Road, +44 1865 272866 (PA) Oxford OX1 3TG, UK Fax: +44 1865 272595
______________________________________________ R-devel@stat.math.ethz.ch mailing list https://stat.ethz.ch/mailman/listinfo/r-devel
