Gentoo Archives: gentoo-amd64

From: Frank Peters <frank.peters@×××××××.net>
To: gentoo-amd64@l.g.o
Subject: Re: [gentoo-amd64] Possible Math Problem - Request For Verification
Date: Sun, 26 Jun 2011 20:34:49
Message-Id: 20110626163317.b9d3749a.frank.peters@comcast.net
In Reply to: Re: [gentoo-amd64] Possible Math Problem - Request For Verification by Mark Knecht
1 On Sun, 26 Jun 2011 12:49:03 -0700
2 Mark Knecht <markknecht@×××××.com> wrote:
3
4 >
5 > Do you have any info about _any_ system passing this 100% successfully?
6 >
7
8 If you look at the log files for various systems that are found in the
9 original source code (http://www.netlib.org/fp) you will see that SunOS
10 passed 100% but SunOS with GCC did not. These results were from 1995.
11
12 > What about that the code itself might have a bug in it?
13
14 The package is not documented at all, but the few references to the ucbtest
15 don't find any fault with it.
16
17 However, this thread has strayed a little from the main theme. I just
18 happened to notice that the cabsd routines failed on my machine and
19 was wondering what could have caused the failure. It turns out that
20 the latest GCC somehow produces bad code.
21
22 The ucbtest itself should not be of much concern. The failures for
23 the trig functions, I assume, are because the error is greater than
24 1 ulp and the upper bound for these errors is usually stated to be
25 1 ulp (although this is not part of the IEEE standard).
26
27 But failures for ceil(), floor(), and cosh() do seem to be of some concern.
28 The file clib_DP.output should be examined to see what actually happened
29 with those functions.
30
31 There are other FP test packages available, although not many, and most
32 of these are without documentation and with messy source code as well.
33 Someone or some organization needs to write a comprehensive, well documented,
34 and neat test package for FP. The problem is that FP is a very abstruse area
35 that few people want to bother with and that even fewer possess the
36 qualifications for.
37
38 Frank Peters