Re: [vox-tech] another gcc question (Branch Delay Slots)
In modern processors you can take this much further
parallel speculative execution and dedicate *HUGE* amounts of chip
space to be able to run *both* sides of the branch operation in parallel
and then forget the results of the side that shouldn't have been run
once you know the outcome of the conditional test... some processors
can even nest this a few levels so that a branch who's both children
also split could in theory not stall the processor.
I'd call this a bit optimistic. Branch prediction is common, even
through multiple branches, intermediate results are stored in shadow
registers and only committed once the branch has been verified as correct.
These shadow registers are not architectually visible. So when a branch
is mispredicted you just clear all the pipelines, don't commit
the related shadow registers and start over.
Not that their are no such machines that exist that follow both branches,
but I bet noone on vox-tech owns one. Hrm, at least the only chip I
know of that has it is the IA64/itanium, last I checked they had sold
2200 of them, which sounded good, till IBM mentioned they had 2000 of
them in a large parallel computer.
Alas branches average 1 in every 7 cycles, cpu pipelines are growing
into the 10-20 cycle range, with 4-8 functional units running you could
easily loose 100's of instructions with a branch mispredict.
The current branch prediction is fairly sophisticated stuff, many cpu's
have more then one algorithm, some put prediction info in the cache line,
some depend on interactions with the compiler to attempt to make the
branchs easier to predict by rearranging the braches to be more similar
to each other. Some strategies include history for the branch, and some
cpu's even implement more then one strategy and then track which is most
accurate for a branch.
I haven't read all of this thread, but it seems like the loop unrolling
typically goes like:
for (i = 0; i 100; i++)
{
do_stuff(i);
}
Crap, if do_stuff isn't much work I spend most of my time doing
a compare, test and jump, so lets unroll:
for (i = 0; i 100; )
{
do_stuff(i); i++;
do_stuff(i); i++;
do_stuff(i); i++;
do_stuff(i); i++;
do_stuff(i); i++;
do_stuff(i); i++;
do_stuff(i); i++;
do_stuff(i); i++;
do_stuff(i); i++;
do_stuff(i); i++;
}
Oh, hrm, what happens if I the number of loops (100) changes? That is
a bit more difficult but we could do something like:
while ((100-i)unroll factor)
for (i = 0; i 100; )
{
do_stuff(i); i++;
do_stuff(i); i++;
do_stuff(i); i++;
do_stuff(i); i++;
do_stuff(i); i++;
do_stuff(i); i++;
do_stuff(i); i++;
do_stuff(i); i++;
do_stuff(i); i++;
do_stuff(i); i++;
}
for (j = i; j 100; j++)
{
do_stuff(j);
}
Then a cool guy named Duff figured out:
j = 0;
count=100;
n = (count + 7) / 8;
switch (count % 8)
{
do
{
case 0:
do_stuff(j);
j++;
case 7:
do_stuff(j);
j++;
case 6:
do_stuff(j);
j++;
case 5:
do_stuff(j);
j++;
case 4:
do_stuff(j);
j++;
case 3:
do_stuff(j);
j++;
case 2:
do_stuff(j);
j++;
case 1:
do_stuff(j);
j++;
}
while (--n 0);
}
Despite it not looking very dense, it's actually quite efficient, and
reasonably flexible. It gracefully handles a dynamic number of loops
that don't have to be a multiple of the unrolling factor.
It's commonly refered to as Duff's device.
Now if you have a smart compiler (that knows how to reorder),
and
the code in question has operations which can be used to fill the
branch delay slots,
and
the chip builders don't want to waste lots of chip space to be
able to speculatively execute
Yes, chip builders like to design simple chips, the $billion question
is how to improve useful work done per cycle with the doubling every
18 month transistor budget.
My current favorite technique that might actually improve this
significantly (work done per cycle hasn't changed much in the last 10
doublings of transistors budgets or so) is SMT, the ability for a cpu to
run multiple threads/processes at the same time during the same cycle.
An alpha (RIP) paper showed a factor of 2.5 in the amount of work done
per cycle with identical cpu resources (except for SMT). Pretty amazing
stuff, especially useful now that quad issue 1.5-2.0 Ghz cpu's are fairly
common and memory is as far away as ever (about 100 ns or so)
Anyways, enough rambling, late.r
--
Bill Broadley
Mathematics/Institute of Theoretical Dynamics
UC Davis
___
vox-tech mailing list
[EMAIL PROTECTED]
http://lists.lugod.org/mailman/listinfo/vox-tech
Re: [vox-tech] regarding config issues wiht Armada e500 (was win98)
-BEGIN PGP SIGNED MESSAGE- Hash: SHA1 It's actualy an Armada V300, but the e500 is an almost identical model On Thursday, February 28 2002 10:14 pm, Steven Peck wrote: Doesn't Compaq systems let you modify them from bootable Compaq CD Roms? I have no idea, I bought the thing used without any CDs. How old is this laptop and have you tried Compaq's support site for this? The have an email response system where you ask a question and they try and get you a response in 24 hours. It has run about 85% accurate on some server issues, which is not bad for email support correct on first answer. I could just call them, Armadas are thier corprate targeted laptops, so the hold is always under 5 minuets. -sp -Original Message- From: [EMAIL PROTECTED] [mailto:[EMAIL PROTECTED]] On Behalf Of Ryan Sent: Thursday, February 28, 2002 8:16 PM To: [EMAIL PROTECTED] Subject: Re: [vox-tech] How slim can Win98 go? On Thursday, February 28 2002 08:14 pm, Mark K. Kim wrote: You could also get a laptop HD from someone. Or an old laptop. Just use the HD, then switch it back when you're done. You don't even need to buy it, just borrow it, if you can find someone with an old laptop laying around. I should have thought of that. Also, many BIOS configuration utilities work under DOS, so you don't necessarily need Windows, but just the DOS (fits on a floppy...) That depends on the utility, though, I suppose. These are Windows utils. -Mark On Thu, 28 Feb 2002, Ryan wrote: -BEGIN PGP SIGNED MESSAGE- Hash: SHA1 hrm, oh, idea! IDE = laptop drive adaptor + old HD Anyone know how much the adaptors cost? Better yet, anyone got one I can borrow? My laptop HD is only a 6 gigger, so I'm a bit cramped to waste 400 MB on a bios configuration utility. On Thursday, February 28 2002 07:33 pm, Mark K. Kim wrote: I've installed a pretty comfortably bare Windows95 that took up about 350MB. Looking at my current Windows98 partition (another comfortably bare installation), it looks like just the Windows part is taking up about 500MB; but this partition's been around for several months so I'm sure there's lots of junk in c:\windows and the registry. -Mark On Thu, 28 Feb 2002, Ryan wrote: -BEGIN PGP SIGNED MESSAGE- Hash: SHA1 I was poking around the net looking for a way to set my laptop's contrast, and I found http://www.informatik.uni-kiel.de/~fst/linux_armada_e500.html . It says that power saving hotkey works in linux, so I could set my cpu to 250 and run for nearly 3 hours. However --- these options need to be set in windows, the bios doesn't have the options in it's menu. :( Anyway, does anyone know what the bare minimum install for windows 98 is? - -- No Microsoft products were used in any way for the creation of this message. PGP Public key at http://mother.com/~ryan/ryan_at_mother_dot_com.asc It is also on the servers: Key ID 0x72177BC7 -BEGIN PGP SIGNATURE- Version: GnuPG v1.0.6 (GNU/Linux) Comment: For info see http://www.gnupg.org iD8DBQE8fsw0Ed9E83IXe8cRAnSpAKDBphWoRnTmjgOoh0oQh0woh1i5dwCgyER6 CBrtKaS/+Z/F3FiPiQRUbn4= =aRcz -END PGP SIGNATURE- ___ vox-tech mailing list [EMAIL PROTECTED] http://lists.lugod.org/mailman/listinfo/vox-tech - -- No Microsoft products were used in any way for the creation of this message. PGP Public key at http://mother.com/~ryan/ryan_at_mother_dot_com.asc It is also on the servers: Key ID 0x72177BC7 -BEGIN PGP SIGNATURE- Version: GnuPG v1.0.6 (GNU/Linux) Comment: For info see http://www.gnupg.org iD8DBQE8fv1DEd9E83IXe8cRAqYhAKCo+u5agZwL5tQUR3obT/ZU45ucEgCgn7gH ARKwE71JXCI+581WWaYYLF4= =m5kr -END PGP SIGNATURE- ___ vox-tech mailing list [EMAIL PROTECTED] http://lists.lugod.org/mailman/listinfo/vox-tech - -- No Microsoft products were used in any way for the creation of this message. PGP Public key at http://mother.com/~ryan/ryan_at_mother_dot_com.asc It is also on the servers: Key ID 0x72177BC7 -BEGIN PGP SIGNATURE- Version: GnuPG v1.0.6 (GNU/Linux) Comment: For info see http://www.gnupg.org iD8DBQE8f5kWEd9E83IXe8cRAn/LAKC/vDYQxU2zUaWKcpTGt0hz/3OGFACggnJC r0/PVCNomHGhOt0WK70SxvY= =tHnU -END PGP SIGNATURE- ___ vox-tech mailing list [EMAIL PROTECTED] http://lists.lugod.org/mailman/listinfo/vox-tech
Re: [vox-tech] How slim can Win98 go?
Another thought is download the trial version of VMWare install Windows, setup the hardware and remove windows / vmware. ___ vox-tech mailing list [EMAIL PROTECTED] http://lists.lugod.org/mailman/listinfo/vox-tech
Re: [vox-tech] How slim can Win98 go?
I bought VMware Express a while ago, but it's of no use, the software would need to have access to the CMOS/BIOS. Another thought is download the trial version of VMWare install Windows, setup the hardware and remove windows / vmware. ___ vox-tech mailing list [EMAIL PROTECTED] http://lists.lugod.org/mailman/listinfo/vox-tech ___ vox-tech mailing list [EMAIL PROTECTED] http://lists.lugod.org/mailman/listinfo/vox-tech
[vox-tech] ip_masq-quake.o
what exactly does this module do? other than opening up the quake ports and forwarding them to the proper internal LAN machine, what do you need a kernel module for in order to serve quake games? pete -- Enron..safe legal abortion..civil liberty..repealing ICBM treaty..deficit.. What's worse? Screwing an intern or screwing an entire country? PGP Fingerprint: B9F1 6CF3 47C4 7CD8 D33E 70A9 A3B9 1945 67EA 951D ___ vox-tech mailing list [EMAIL PROTECTED] http://lists.lugod.org/mailman/listinfo/vox-tech
Re: [vox-tech] ip_masq-quake.o
The definitive site... well it was. Now it looks like you'll have to do some wading http://www.tsmservices.com/masq/cfm/main.cfm Oh wait, it seems to be here: http://www.e-infomax.com/ipmasq/ IIRC it so that the appropriate connections are made and MASQ'd properly if you are playing Quake, or any number f other games that use essentially the same port settings behind a NAT'd network. If you host a Quake server, you have to open up ports differently. -sp On Fri, 01 March 2002, Peter Jay Salzman wrote: what exactly does this module do? other than opening up the quake ports and forwarding them to the proper internal LAN machine, what do you need a kernel module for in order to serve quake games? pete -- Enron..safe legal abortion..civil liberty..repealing ICBM treaty..deficit.. What's worse? Screwing an intern or screwing an entire country? PGP Fingerprint: B9F1 6CF3 47C4 7CD8 D33E 70A9 A3B9 1945 67EA 951D ___ vox-tech mailing list [EMAIL PROTECTED] http://lists.lugod.org/mailman/listinfo/vox-tech ___ vox-tech mailing list [EMAIL PROTECTED] http://lists.lugod.org/mailman/listinfo/vox-tech
Re: [vox-tech] ip_masq-quake.o
On Fri, Mar 01, 2002 at 12:34:08PM -0800, Peter Jay Salzman wrote: what exactly does this module do? other than opening up the quake ports and forwarding them to the proper internal LAN machine, what do you need a kernel module for in order to serve quake games? I don't think this is intended to help with serving It seems to handle the forwarding of Quake packets to clients behind a masq box, so that when a client behind the masq box connects to a server, UDP packets from the server are sent back to the client Kernel 24 seems to do this automagically I've never loaded any Quake-specific kernel modules on my 24-based masq box, and I've never had any problems with Quake (Q3|Quakeworld) playing -- Samuel Merritt PGP key is at http://wwwcsifcsucdavisedu/~merritt/snmerrittasc Information about PGP can be found at http://wwwmindspringcom/~aegreene/pgp/ msg01663/pgp0.pgp Description: PGP signature
Re: [vox-tech] How slim can Win98 go?
Well, I am not sure If I want to part with it, but I could let you borrow it for awhile. It is actually a 8.45mm drive instead of the 12mm or 19mm thickness that some older laptops had. I did see 2GB ($39.99) and 4GB ($49.99) laptop drives on weekly special at halted.com Ryan wrote: I have a lot of junk. I have about 1.8 gigs of free space. The drive would be great, thanks. Will you need it back (I'm fine either way)? ___ vox-tech mailing list [EMAIL PROTECTED] http://lists.lugod.org/mailman/listinfo/vox-tech
Re: [vox-tech] ip_masq-quake.o
doug, what exactly do you mean by related ip traffic? pete begin [EMAIL PROTECTED] [EMAIL PROTECTED] IIRC, ipchains and ipfw didn't handle related ip traffic. It was one of the most annoying features of that firewalling code. You had to load other modules so that related ip traffic could follow the same rules. With the masq modules you could define one set of rules for say ftp and then realted ip traf could be handled. I remember having alot of trouble with FTP until I RTFM and loaded the right modules. -doug ___ vox-tech mailing list [EMAIL PROTECTED] http://lists.lugod.org/mailman/listinfo/vox-tech
Re: [vox-tech] ip_masq-quake.o
i just stumbled across something which seems to support this -- it said that the kernel module is used when 2 masqueraded machines want to play an internet game at the same time. pete begin Samuel Merritt [EMAIL PROTECTED] On Fri, Mar 01, 2002 at 12:34:08PM -0800, Peter Jay Salzman wrote: what exactly does this module do? other than opening up the quake ports and forwarding them to the proper internal LAN machine, what do you need a kernel module for in order to serve quake games? I don't think this is intended to help with serving. It seems to handle the forwarding of Quake packets to clients behind a masq box, so that when a client behind the masq box connects to a server, UDP packets from the server are sent back to the client. Kernel 2.4 seems to do this automagically. I've never loaded any Quake-specific kernel modules on my 2.4-based masq box, and I've never had any problems with Quake (Q3|Quakeworld) playing. -- Samuel Merritt PGP key is at http://wwwcsif.cs.ucdavis.edu/~merritt/snmerritt.asc Information about PGP can be found at http://www.mindspring.com/~aegreene/pgp/ -- Enron..safe legal abortion..civil liberty..repealing ICBM treaty..deficit.. What's worse? Screwing an intern or screwing an entire country? PGP Fingerprint: B9F1 6CF3 47C4 7CD8 D33E 70A9 A3B9 1945 67EA 951D ___ vox-tech mailing list [EMAIL PROTECTED] http://lists.lugod.org/mailman/listinfo/vox-tech
Re: [vox-tech] How is lugod.org put together?
On Thu, Feb 28, 2002 at 08:55:52PM -0800, Peter Jay Salzman wrote: although -- wait. isn't there a dotfile.org already? dotfiles.org. Seem to have gone away and been replaced with a big 'coming soon' page. :( Domain Name: DOTFILES.ORG Administrative Contact, Technical Contact, Billing Contact: Momoi, Yasunari (YM233) [EMAIL PROTECTED] BUG.ORG 2-27-1-601 Ayase Adachi-ku Tokyo 120-0005 JP +81-90-7731-5086 (FAX) none Record last updated on 08-Jan-2002. Record expires on 01-Mar-2004. Record created on 01-Mar-2000. Database last updated on 1-Mar-2002 07:48:00 EST. -bill! ___ vox-tech mailing list [EMAIL PROTECTED] http://lists.lugod.org/mailman/listinfo/vox-tech
