NEON code optimization

Hello,

Anyone having idea about how to optimize code made using NEON
instructions?

Thanks,
Seema

Hello,

Anyone having idea about how to optimize code made using NEON
instructions?

http://www.opensdr.com/node/13

There is a brief introduction here. Most of the actual NEON is based
on suggestions from Mans.

Philip

Hi, Seema:

Did you hear the OpenMAX project ? you might reference OpenMAX and use
this library directly.

Hi, I am using the angstrom source code and code sourcery as a IDE for
development. I have developed an algorithm In which I am using NEON
and ARM both instructions. Can you please tell me how can I use this
library in that?

Thanks,
Seema

Seema,
1.First go through the documentation supplied with the compiler and
understand the neon instructions and thier capabilities.
2.Vectorize the C Code (Making the C Code Neon friendly so that you
can use vector instructions)
3.Code the computation intensive functions in Neon assly.
4.Profile and see the results.
5.You are done

What code you wanna optimise?
Regards,
Sanjeev Kumar Verma

If you're targeting angstrom make sure you compile the resulting binaries NOT with codesourcery, but with the angstrom toolchain to avoid nasty surprises.

Seema,
You need to modify the corresponding make files to include your
library.
Regards,
Sanjeev Kumar Verma

Ok thanks,

I have converted my code in NEON assembly. I have done profiling too.
But I still want to optimize it. May be it will be possible if I get
the events count of cache hit and cache miss.

Thanks,
seema

Seema,
You can not measure the cache hits and misses directly. But ceratainly
you need to study the cache behaviour of the processor you are using
and then modify your code to maximise the cache hits.
However there is an indirect way to measure the cache hits. Write some
simple read word and write word kind of code and then see how many
cycles it is taking. If it takes more than you expect, it is due to
the cahe misses.If it takes as per your expectation than it means that
cache hits are happening.

Ragards,
Sanjeev Kumar Verma