Subject: RE: [linux-audio-dev] MMX, SSE, SSE2, 3DNOW
From: Bob Colwell (bob.colwell_AT_attbi.com)
Date: Mon Feb 18 2002 - 20:32:41 EET
"wacky sub-architectures?" I beg to differ. :-)
There are really only 4, and I think I can argue them down to two, at
least for Intel boxes.
In the beginning was the Intel 386, and its pipelined younger brother
the 486. And between Intel and AMD, a lot of those chips got sold and
most of the world loved them for it.
Then came the Pentium, and with its 2nd generation, the MMX integer
SIMD instructions. Pentium II also included MMX.
Next came the Pentium III, which had MMX and the single
precision version of MMX, called SSE. I'm leaving out the Pentium Pro
because those chips shipped mainly into servers.
Then came the Pentium 4, with MMX, SSE, and the double-precision version
of SSE, called SSE2.
This progression occurred because of the algorithm being followed at
Intel: put in whatever fits, and make sure the memory bandwidth to
support it is there.
I'll be really surprised if enough interesting sound work was being
done on 386's and 486's that we need to even remember those. In fact,
are there really enough Pentium boxes remaining in the field to care
about them and their MMX-only capability?
I'm not so sure about Pentium II machines, which were shipped as high
as 400 MHz, and might be still fast enough and widely enough used that
we still care about them in this context.
To look at all this another way, what's the problem if we restrict our
attention here to just Pentium III and newer designs? If the new
instruction sets are enough of a win to ever use, then they're also
enough of a win to entice people to buy up. They have to, sooner or
later.
-Bob Colwell
-----Original Message-----
From: owner-linux-audio-dev_AT_music.columbia.edu
[mailto:owner-linux-audio-dev_AT_music.columbia.edu]On Behalf Of Steve
Harris
Sent: Monday, February 18, 2002 7:23 AM
To: linux-audio-dev_AT_music.columbia.edu
Subject: Re: [linux-audio-dev] MMX, SSE, SSE2, 3DNOW
On Sat, Feb 16, 2002 at 09:09:24 +0100, Alexander Ehlert wrote:
> Hi,
>
> AFAIK there are no flags whatsoever to indicate which processor a plugin
> might use. So if someone wants to hack a plugin that uses SSE
> instructions and someone else tries to use that on a host without SSE
> support -> crash. So wouldn't it be good to add some architecture flags,
> that could be queried by the host?
So far, I have only distributed binaries in RPM only, which has
architecture dependencies. Actually they are wrongly labelled as i386,
they are really i686.
I don't think the plugin should pick a code block based on some detection
code, that would make the plugins image larger, and therefore less
efficient. There are far too many wacky sub-architectures of x86's to
include SIMD instructions for all of them in one binary.
I have plyed with multi-format binaries (see mail later today), and it is
tricky.
I think the right thing to do is to take the same approach as with normal
libraries and binaries, and label the packages.
- Steve
This archive was generated by hypermail 2b28 : Mon Feb 18 2002 - 20:25:08 EET