hand crafted SSE2 matrix&vector ?

Hi all,

anyone thought of, implemented or analyzed the potential of hand crafted versions of core vector/matrix calcs using SSE2 instructions on Intel&AMD CPUs ?
(apart from "-march=pentium4 -mfpmath=sse")

-Peter

···

--
pab-opto, Freiburg, Germany, http://www.pab-opto.de
[see web page to check digital email signature]

Good idea.

I recommend some profiling ahead of time to determine where to concentrate in the code. I suspect that the vector and matrix routines, although inviting, don't use a substantial fraction of the time in most renderings. If short vectors don't kill you on set-up, the color manipulation macros might be a better starting point, or perhaps something could be done to accelerate the expression evaluation code in the common/cal* routines. I know these can be a bottleneck in a lot of scenes.

In general, a good profiling of the code on a variety of scenes would be very helpful to anyone wishing to work on tuning Radiance in the future.

My humming desktop also compels me to mention that Apple has some handy C libraries for vector and matrix operations on the G4 and G5 processors. (I haven't tried them yet myself, I'm ashamed to say.)

-Greg

···

From: Peter Apian-Bennewitz <[email protected]>
Date: October 1, 2004 9:46:29 AM PDT

Hi all,

anyone thought of, implemented or analyzed the potential of hand crafted versions of core vector/matrix calcs using SSE2 instructions on Intel&AMD CPUs ?
(apart from "-march=pentium4 -mfpmath=sse")

-Peter

I am currently interested in camera
lens/glare/flare/lenticular halo/depth of field/motion
blurr etc. post-processing of Radiance images and to
that aim I joined blitz++
(http://www.oonumerics.org/blitz ) and FFTW (
[email protected]) groups.

I think that code and research available from these
two sites is very promising.

Unfortunately, in my experience so far, it seems that
these people suffer from the
parochial/boy-scouts/Phi-Gamma-Whatever/golf-buddies
mentality and despite all the pretense and claims to
the contrary, they typically don't answer emails or
demonstrate any capability to communicate with the
'outsiders'.

Tony

···

--- Peter Apian-Bennewitz <[email protected]> wrote:

Hi all,

anyone thought of, implemented or analyzed the
potential of hand crafted
versions of core vector/matrix calcs using SSE2
instructions on
Intel&AMD CPUs ?
(apart from "-march=pentium4 -mfpmath=sse")

-Peter

__________________________________
Do you Yahoo!?
Yahoo! Mail - 50x more storage than other providers!
http://promotions.yahoo.com/new_mail