quote from anger pOpt doc

All these packing modes are available on the latest microprocessors from Intel and AMD,
except for the 3DNow mode, which is available only on AMD processors.

which AMD processors support SSe and SSE2??

Karl
Posted on 2004-06-02 20:26:09 by klumsy
The oringinal Athlon and Thunderbird core don't have SSE support. All Athlon's since Palomino support SSE. No Athlon supports SSE2 currently, but Opteron's support SSE2.

This is easily found on the web, and I may have misstated something, but I do own a Thunderbird (no SSE) and Barton (SSE) core Athlon.
Posted on 2004-06-02 23:48:54 by bitRAKE
Athlon64 also supports SSE2 (then again, it is based on Opteron).
No AMD processor supports SSE3 at this time though.
Posted on 2004-06-03 07:25:02 by Scali
What programs support SSE3?
Posted on 2004-06-03 09:57:01 by bitRAKE
Are you trying to find an excuse not to need SSE3?
Then you fail, since at the time of introduction, no program supported MMX, SSE or SSE2 either, and look what happened!
Now stop acting like a useless AMD zealot, this is the second time now, and it is starting to get silly.
Posted on 2004-06-03 14:44:44 by Scali
Just curious - nothing more. SSE3 has been out for some time and I've heard of no applications using it. Does anyone know of the actual speed increases gained by use of the instructions.

http://www.intel.com/technology/itj/2004/volume08issue01/art01_microarchitecture/p06_sse.htm
http://www.xbitlabs.com/articles/cpu/display/prescott_10.html

I can find links on the instructions, but the benefit of so few new instructions is not very clear at this time - not that it could be.
Posted on 2004-06-03 16:32:39 by bitRAKE
i never knew of the existence of SSE3, what do all p4 processors support it or what processors do?

i looked at hte first example? does SSE and SSE2 not have add and sub with saturation expressions like MMX does? (since it says it is added in SSE3) or am i barking up the wrong assumption tree
Posted on 2004-06-03 18:13:24 by klumsy
for Intel CPU's: SSE3 (SSE2 (SSE (MMX) ) )
Posted on 2004-06-03 18:42:31 by bitRAKE
ho humm. What exactly does SSE3 add? MMX added stuff a lot of people could use (after a while), SSE/2 took a while to use but has been incorporated into varioues video codecs. What does SSE3 add?
Posted on 2004-06-03 19:17:08 by f0dder
SSE3 has been out for some time and I've heard of no applications using it.


"Some time"? They have only been in Prescott, and that was released in February. Hardly any software has had the chance to go through the next update cycle yet.

Does anyone know of the actual speed increases gained by use of the instructions.


As you should know, there is no net speed gain from specialist instructions such as MMX/SSE/SSE2. It all depends on where and how you can deploy them.

does SSE and SSE2 not have add and sub with saturation expressions like MMX does?


No. I suppose it doesn't make that much sense with floats as it does with integers, so they saved it for later? :)

What exactly does SSE3 add?


It gives you 'horizontal' add/sub operations which make operations on transposed vectors/matrices more efficient, and also things such as dotproducts.
There's also a combined add/sub operation, which will no doubt be useful in DCT/DFT/FFT operations.
There is some stuff to hint HTT so that it is even more efficient...
And there is a new fistp instruction which always truncates to int in the C-tradition, so no need to screw around with the control word anymore (and why did it take them THIS long to implement it? 68k has had it for decades).

So there's some very useful stuff there. I suppose AMD needs to implement HTT first, before they can implement SSE3 completely... or they have to convert the HTT-related stuff to nops perhaps.
Posted on 2004-06-04 02:17:24 by Scali
With SSE2 all (i think) mmx instructions can be used on xmm registers so you can have



paddusb xmm0,xmm1
psubusb xmm1,xmm2


and stuff.
Posted on 2004-06-04 07:07:57 by stormix
but those mmx comands with the xmm registers how does it work? is it taking the xmm value as a float, or an integer? i.e you couldn't just load say ARGB values into it, and use it to process them?
Posted on 2004-06-06 00:14:40 by klumsy
They work on ints, just like MMX. The registers are just twice as wide.
Posted on 2004-06-06 04:27:04 by Scali
so XMM can act as ints rather than floating point ? in some of the PentOpt document it said something about you can't just use XMM for any data.. it had to be a vcalid floating point number or something?
Posted on 2004-06-07 15:45:37 by klumsy
well, when you use the *MMX* opcodes with *XMM* registers, afaik it just adds one of those size prefix bytes...
Posted on 2004-06-07 15:54:42 by f0dder