floating point ops in bf533

Hi all,

I have a problem, i will really really appreciate if you can give some comments or suggestions.

I am implementing the AMR codec in BF533 in real-time. Since the codec processes speech frame of length 20ms, I am allocating time<20ms to perform encoding such that after this, the next frame is almost ready.

my problem is this, when i checked the profiling when i run the encoder for the first frame, the encoder consumed 200ms! and 50% is due to ___float32_sub and ___float32_mul which I assumed are floating point operations.

do you have any idea how to lessen the burden of the BF533 when it comes to the floating point operation? i will really really appreciate any suggestions that you will give. God bless.

regards,
mac Start a New Thread

The blackfins are architected as fixed-point processors. So,
floating-point is really slow. I'd suggest doing your math in fixed
point, or moving to a different processor. I think ADI's Sharc family of
processors are floating point.

David

-----Original Message-----
From: a... [mailto:a...] On Behalf Of
maccordel
Sent: Thursday, July 15, 2010 6:37 AM
To: a...
Subject: [adsp] floating point ops in bf533
Hi all,

I have a problem, i will really really appreciate if you can give some
comments or suggestions.

I am implementing the AMR codec in BF533 in real-time. Since the codec
processes speech frame of length 20ms, I am allocating time<20ms to
perform encoding such that after this, the next frame is almost ready.

my problem is this, when i checked the profiling when i run the encoder
for the first frame, the encoder consumed 200ms! and 50% is due to
___float32_sub and ___float32_mul which I assumed are floating point
operations.

do you have any idea how to lessen the burden of the BF533 when it comes
to the floating point operation? i will really really appreciate any
suggestions that you will give. God bless.

regards,
mac Start a New Thread

You can also try to use a simplified floating point format as described here:
http://www.analog.com/static/imported-files/application_notes/EE.185.Rev.4.08.07.pdf

Alexander

Hi all,
>
>I have a problem, i will really really appreciate if you can give some comments or suggestions.
>
>I am implementing the AMR codec in BF533 in real-time. Since the codec processes speech frame of length 20ms, I am allocating time Start a New Thread

Hi,
Mostly ppl use fixed pt processors for video processing and floating pt
processors for Audio. So I suggest you to use a floating pt processors like
SHARC. Or in case if it is mandate for you to use blackfin, you can try out
fixed pt fractional notations.

On Mon, Jul 19, 2010 at 3:49 AM, wrote:

> You can also try to use a simplified floating point format as described
> here:
>
> http://www.analog.com/static/imported-files/application_notes/EE.185.Rev.4.08.07.pdf
>
> Alexander
> Hi all,
> >
> >I have a problem, i will really really appreciate if you can give some
> comments or suggestions.
> >
> >I am implementing the AMR codec in BF533 in real-time. Since the codec
> processes speech frame of length 20ms, I am allocating time
>
>
> Start a New Thread