On Thu, Mar 15, 2001 at 03:25:46PM +0100, Francisco Gómez-Molinero wrote:
> We are working with the SA1110 for image processing applications. We
> have come across the performance library of Intel. Although the
> primitives in the library give good performance, they are very much
> oriented to implementing image decompression. Is anybody out there
> working in image compression primitives for the SA and willing to share
> experiences? In particular, is any forward DCT application available
> somewhere?
We have discussed the IPP (Intel Performance Primitives) library on the
Intel StrongARM Linux summit in New York, and according to a lot of
people the IPP library is potentionally useful, but comes with the
wrong license that makes it incompatible with GPL'ed programs. I think
we convinced the Intel handhelds division (the guys behind the
StrongARM and XScale) to release the IPP library with source code under
a more open license, but so far we have failed to convince the people
behind the library itself.
Of course the usual Open Source arguments apply to this library, but
there are some more:
- Intel's core business is to sell CPUs. Selling more CPUs means more
money for Intel. If there is software that makes a CPU more useful,
give it away.
- Giving away the library is good for Intel's PR. Yes, there is a risk
that the library will be ported to non-Intel CPUs, but there is still
"Intel" in the name of the library.
- You can't trust the results from closed source software. I've seen
this more than once: we upgraded a closed source image processing
package, and couldn't reproduce our results. Quite an embarrassing
experience from a scientific point of view. The IPP library is closed
source, so it can't be used in *any* academic environment.
- To be useful in an academic environment, I also miss implementation
details. For exmple, what's the noise that each function adds? How is
it implemented? There are quite some ways to implement a certain
algorithm, and there is no way for me to find out if Intel used the
smart way. For an example, read about "random number generators" in
"Numerical recipes in C".
- Intel claims to have the smartest algorithm designers on the planet.
We think we can prove them wrong with our FFT implementation. I am
from a video processing group and we have quite some good algorithm
experts as well. I'm sure your group also has some. Intel should
*use* the knowledge from those people by giving away the IPP source
and let people contribute to it. For a business case, ask Compaq
about running Linux on the iPaq.
- In some cases you can gain quite some speed by folding some
algorithms together into a single function. For example: in RF
channel coding, where the combination of the error correction encoder
with the first stage of the FFT can pretty much make the ECC function
"free" as loads, stores and cache reloads are eliminated. Same for a
H.263 decoder: folding the YUV-->RGB mapping and RGB-->palette
mapping into a single function made the decoder four times as fast
because we could *calculate* the correct palette entry instead of
having to use two lookup tables (that trashed the cache).
Wookey, Nicolas Pitre and me tried to convice Stewart Taylor (one of
the IPP head architects) after his presentation on Linux World Expo,
and we are currently trying to get in touch with the IPP people
directly, but they are quite hard to reach.
Anyway, to answer your original question: I am not aware that there is
an optimised forward DCT library. If the source was available, you
could have used the IDCT function as a skeleton for a DCT function and
contributed your code back to Intel. Now you have to write it yourself
from scratch. I'd call this a loose-loose situation.
Erik
--
J.A.K. (Erik) Mouw, Information and Communication Theory Group, Department
of Electrical Engineering, Faculty of Information Technology and Systems,
Delft University of Technology, PO BOX 5031, 2600 GA Delft, The Netherlands
Phone: +31-15-2783635 Fax: +31-15-2781843 Email: [EMAIL PROTECTED]
WWW: http://www-ict.its.tudelft.nl/~erik/
_______________________________________________
http://lists.arm.linux.org.uk/mailman/listinfo/linux-arm
Please visit the above address for information on this list.
