The VXL Project

Short answer - this is entirely non-trivial - using BLAS would be better.


Long answer - As I understand it:
MMX is a means of performing SIMD on narrow integers (either 8 bit with 4
fold parallelism, or 16 bit with a 2 parallelism.)

Most of VNL is bound by computations on the type double. You would therefore
need to at the very least ensure that the code you are interested in runs in
terms of fixed-point real representations.
This is difficult
	You have to ensure that your real numbers don't over or underflow (You can
only represent numbers from x to x*2^(-n) where n is 8 or 16, and x is the
largest number you want to represent. Many interesting operations take
vastly more dynamic range than that.
      You would need to rewrite the bits of VXL that aren't templated so
that they used your integer format.


Alternatively there is another (its not MMX, called something different I
believe) method that does SIMD on pairs of single-precision floats. Both the
above points still apply. The dynamic range of floats is considered too low
for many interesting operations (SVD, etc.)

I would have thought that a faster, more useful thing to do would be to
integrate BLAS into VNL so that we can use the properly optimised processor
specific BLAS routines. This could give a speed of more that 2.

Ian.

> -----Original Message-----
> From: Dominique [mailto:domi@...]
> Sent: Wednesday, June 26, 2002 11:56 AM
> To: Vxl-users@...
> Subject: [Vxl-users] misc questions
>
>
> Hi,
>
> Is it possible - and if yes: how - to compile vxl's numerics taking
> adventage of MMX (pentium/pc). I just thought about it compiling lame
> (mp3 enc.) - if "nasm" (x86asembler) is present on the system it is
> possible to build also MMX executable version that is more
> than twice as
> fast as the normal one.
>
> any ideas to start with?
>
> thanks
> Domi
>
> --
> --------------------------------------------------------------------
> Dominik Szczerba, Dr.             Phone: +41 1 632 66 68
> Computer Vision Lab, ETH          Fax:   +41 1 632 11 99
> Gloriastr. 35                     email: domi@...
> CH-8092 Zurich                    http://www.vision.ee.ethz.ch/~domi
> --------------------------------------------------------------------
>
>
>
>
> -------------------------------------------------------
> This sf.net email is sponsored by: Jabber Inc.
> Don't miss the IM event of the season | Special offer for
> OSDN members!
> JabConf 2002, Aug. 20-22, Keystone, CO http://www.jabberconf.com/osdn
> _______________________________________________
> Vxl-users mailing list
> Vxl-users@...
> https://lists.sourceforge.net/lists/listinfo/vxl-users
>

> Short answer - this is entirely non-trivial - using BLAS would be better.

Thank you for dedailed answer.

As to BLAS - I studied its docs a bit and they say they provide only
general non-optimized stuff. Like in lapac. They encourage to use ATLAS
instead, BLAS succesor as I understood, which should be better to
generate optimized stuff for given architecture. Altas is linked to from
BLAS FAQ page on netlib.

bye
Domi

> I would have thought that a faster, more useful thing to do would be to
> integrate BLAS into VNL so that we can use the properly optimised processor
> specific BLAS routines. This could give a speed of more that 2.
>
> Ian.