[<prev] [next>] [thread-next>] [day] [month] [year] [list]
Date: Sun, 16 Aug 2015 18:44:09 +0300
From: Solar Designer <solar@...nwall.com>
To: john-dev@...ts.openwall.com
Subject: relbench over/underflow when calculating geometric mean
Frank, magnum -
The attached patch allows relbench to calculate the geometric mean even
when there's a large number of significantly different benchmark
results. Without this patch, I was getting "inf" for $mr and (luckily)
a crash on log() on line 266 when comparing OpenMP benchmarks for all of
jumbo's formats on super for 1 vs. 15 threads. This is currently 403
total benchmarks and each may be sped up by ~15x (ideally), resulting in
the naive implementation's intermediate value exceeding the maximum
value for "double". (And comparing these benchmarks the other way, we'd
similarly have an underflow.)
I think we should commit this into jumbo for now, even though the code
comes from core. There are by far not enough formats in core to trigger
the problem in there.
Alexander
View attachment "relbench-geoflow.diff" of type "text/plain" (1019 bytes)
Powered by blists - more mailing lists
Confused about mailing lists and their use?
Read about mailing lists on Wikipedia
and check out these
guidelines on proper formatting of your messages.
