Re: you can't bash Microsoft enough
- From: John Larkin <jjlarkin@xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx>
- Date: Fri, 04 May 2007 08:56:38 -0700
On 4 May 2007 02:04:25 -0700, websnarf@xxxxxxxxx wrote:
On May 3, 2:11 pm, John Larkin
<jjlar...@xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx> wrote:
On 3 May 2007 12:30:29 -0700, websn...@xxxxxxxxx wrote:
On May 3, 10:00 am, John Larkin
<jjlar...@xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx> wrote:
On Thu, 03 May 2007 16:38:02 +0000, Mike Monett <N...@xxxxxxxxx> wrote:
John Larkin <jjlar...@xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx> wrote:
> I sometimes use a little math/ascii conversion package I wrote for
> the 68K. The math format is 32.32, a longword of signed integer
> plus a longword of fraction. That's adequate for near any
> real-world variables. It's blindingly fast compared to floats (no
> normalizing!) and conversion into/out of integers is
> instantaneous. All the math functions saturate, so most of the
> time it's safe to ignore exceptions, including division by 0.
> John
If you haven't come across it yet,Paul Hsiehhas several pages on
integer math for the x86 that might be useful:
"Integer multiplying by constants"
the x86 ISA also includes lea, sub and mov (for shuffling partial
results amongst other registers.) Instructions like sub and mov are
self explanatory, but lea is a very curious and powerful instruction
that gives some multiregister linear combinations as well as some
additional odd multiplies. The x86's "lea" instruction allows you to
compute:
r0 := r1 * {0,1} + r2 * {0,1,2,4,8}
for any 3 general purpose 32 bit registers, in a single clock.
"sub" (subtract) and "mov" (copy value to/from register) are self
explanatory. So using these instructions, we see that 181 =
4*(5*9)+1 which can be encoded as: b = x*5; b = b*9; return x + b*4;
http://www.azillionmonkeys.com/qed/amult.html
"Pentium Optimal x86 Constant Multiplies"
http://www.azillionmonkeys.com/qed/amultl2.html(235k)
"Integer division and modulus by constants"
http://www.azillionmonkeys.com/qed/adiv.html
"Square Roots"
http://www.azillionmonkeys.com/qed/sqroot.html
Regards,
Mike Monett
The 8008 (I think they call the latest version "Core Duo" or
something) is a grotesque hack that needs all the help it can get. The
68332 has 32x32 -> 64 signed or unsigned multiply and 64/32 divide
with remainder, all in hardware.
Hmmm ... are you aware that these puny "Core Duos" are able to do
64x64->128 signed/unsigned multiply in hardware (in about 4 clocks) as
well as a 128/64 divide?
Hmmm... are you aware that their heatsink is bigger than many of our
entire products?
Are you aware that you are changing the subject in order hide the fact
that you were pointing out a *WEAKNESS* in your favorite 68332 CPU not
a strength? Yeah, those extra thousands of Mhz draws more heat.
[...] Our latest 4-channel digital delay generator uses
about 4 watts total, and the FPGA gets warmer than the CPU.
If you wish to decrease the heat draw from a modern CPU, can decrease
the voltage and under-clock the damn thing. Some embedded versions of
the x86 such as AMD's geode processor (which has capabilities more
similar to the 68332 you were talking about)
Geode is still an x86, and National sold it off to AMD just to get out
of the business. It's been tarted up to the extent that it's pretty
much unsuited to low-power, instrumentation sorts of embedded apps.
also have very low heat
draw, while running at several hundred Mhz (I don't know the details,
you can look them up.)
Umm, aren't details what electronic design is all about? The
bleeding-edge nanometer chips, like the current-gen Intel stuff, have
serious static power dissipation because of gate leakage and such, so
reducing the clock rate still leaves you with 10's of watts.
The problem about using an x86 in an embedded product is that, aside
from the grotesque architecture, the product lifetime is typically a
serious multiple of the processor lifetime. Intel has more than once
introduced a mobile or embedded CPU, promoted design-ins, and canceled
the chip while it was still in the sampling stage. Intel also promoted
"Xscale" (actually their version of Arm) and then lost interest and
dumped it. Motorola manufactured the original depletion-load 6800 for
about 20 years. The 68332 is over 10 years old and will likely be
around long after other chips have gone through multiple
intro-obsolete cycles. Sure, the 8052 will be around forever, but I
consider that a crime against humanity.
The bottom line is that what we're doing works. THAT is what
electronic design is about.
Take the Windows off that 3 Ghz PC of yours, put VxWorks on it and it
can service *MILLIONS* of interrupts per second with 100% reliability.
And then jam it into a product the size of a cell phone?
John
.
- References:
- Re: you can't bash Microsoft enough
- From: MooseFET
- Re: you can't bash Microsoft enough
- From: John Larkin
- Re: you can't bash Microsoft enough
- From: MooseFET
- Re: you can't bash Microsoft enough
- From: John Larkin
- Re: you can't bash Microsoft enough
- From: Mike Monett
- Re: you can't bash Microsoft enough
- From: John Larkin
- Re: you can't bash Microsoft enough
- From: websnarf
- Re: you can't bash Microsoft enough
- From: John Larkin
- Re: you can't bash Microsoft enough
- From: websnarf
- Re: you can't bash Microsoft enough
- Prev by Date: Re: Semi-OT: How to throttle a AMD Turion?
- Next by Date: Re: Coil EMF volt vs current
- Previous by thread: Re: you can't bash Microsoft enough
- Next by thread: Re: you can't bash Microsoft enough
- Index(es):
Relevant Pages
|
