Re-sending, because the message size limit is 40k.
This is part 1 of 2, with the Core Duo/OSX data
-----Original Message-----
From: Cornea, Marius
Sent: Wednesday, July 19, 2006 10:56 AM
To: r754@xxxxxxxxxxxxxx; stds-754@xxxxxxxx; stds-754@xxxxxxxxxxx
Cc: Cornea, Marius
Subject: Regarding Items 15 and 17 on the agenda today (decimal
floating-point)
All,
Last month I sent some data to this list, comparing the Intel
implementation of the BID-based decimal floating-point library with the
decNumber package. (The Intel library uses the IEEE 754R binary encoding
for decimal floating-point values, and decNumber uses the IEEE 754R
decimal encoding).
We tried to address three observations that were made related to our
posting of performance data:
1) that we did not use the best decNumber library available
2) that we did not have any results for 32-bit systems
3) that we did not provide the code for independent evaluation
To address the first issue, we re-ran our tests using the latest release
of decNumber, version 3.32. The attached results were obtained using
this version.
For the second one, we ran the tests on a 32-bit Apple system with an
Intel Core(tm) Duo processor (the results are very good, but we must
note that availability of 64-bit operations in hardware will increase
the BID library performance both in absolute terms and relative to
decNumber). One of the attached tables contains the results we obtained
on the Apple Core Duo system. The second table contains results obtained
on a new Intel Xeon 5100 series system (which has 64-bit integer support
in hardware).
To address the third observation, a couple days ago we were able to
package the library ("as is", pre-beta v0.1) with a limited license
agreement (EULA) and to make it available for evaluation to two
different teams. I hope they will be able to present some data in the
754R committee meeting today.
Once again, our point is that the binary encoding format (BID) is an
extremely valuable option for users of the IEEE 754R decimal
floating-point arithmetic, and should be part of the final draft.
Thanks,
Marius Cornea
Intel Corporation
Attachment:
DecimalFpData070506_CoreDuo_OSX.pdf
Description: DecimalFpData070506_CoreDuo_OSX.pdf