On Fri, Feb 26, 1999 at 08:17:46AM -0800, John R Pierce wrote:
> probably not. While I haven't actually seen the tech ref on the KNI
> instructions they are likely to be single precision vector/matrix operations
> oriented towards 3D, probably operating on two 'singles' at a time.. The LL
> test would benefit most from operations that used larger word sizes, perhaps a
> SIMD 64 bit integer multiply-accumulate.
There are some documents on the KNI instructions at
http://www.tbcnet.com/~clive/knirough.html
Basically, KNI (also known as SSI, Streaming SIMD Instructions) allow
the application to execute one instruction (divide or add, for
example) on 4 single-precision floats simultaneously. There are some
instructions which can pertain to integer values, like XOR, AND, and
the like, but for the most part these are aimed at making 3D games go
faster.
A KNI-enhanced version of DES or RC5 will fly, I think, because of the
ability to XOR 128 bits at once. Also, the PSHUFW instruction allows
you to transpose words in one cycle, which is great for RC5 if I
understand correctly.
Also interesting are the uncached instructions, which load and store
directly from memory without polluting the L1 or L2 cache. (There's
no point in using the cache if you're writing 4MB of screen data,
because you're not going to go back and read any of that data before
it gets blown out of the cache anyways.)
I don't see much application to Mersenne searching, though.
-andy
--
Andy Isaacson [EMAIL PROTECTED] [EMAIL PROTECTED] Fight Spam, join CAUCE:
http://www.csl.mtu.edu/~adisaacs/ http://www.cauce.org/
________________________________________________________________
Unsubscribe & list info -- http://www.scruz.net/~luke/signup.htm
