Thank you very much, all of you.
Vielen Dank an euch alle. :)
Reply by Andrew Reilly●November 16, 20112011-11-16
On Mon, 14 Nov 2011 07:22:46 -0600, Caspian wrote:
> I'm currently working on an optimized version of FFT for ARM NEON. (FHT,
> to be exact.)
>
> I'm neither a DSP specialist nor mathematician, and it was very hard for
> me to understand the way FFT works.
>
> Now it's about 50% done, and I have my doubt about this : "What's the
> use of this? What does it bring when my FFT runs 20times faster than
> some unnamed opensource ones on handhelds?"
>
> I really don't see any point int this. Maybe because I know too little
> about DSP.
>
> Could someone enlighten me? It would be really great to be motivated
> again.
Many of the machines with ARM chips in them spend a lot of their time
playing (or, sometimes, recording) music, or watching videos. Most of
the popular audio codecs operate in the frequency domain (usually MDCT or
CQMF or both), so decoding those bitstreams always finishes with a
frequency->time domain transform. Most such transforms can be expressed
as twiddles around FFT routines, so if you have a fast FFT you also have
a fast MDCT or CQMF or whatever. Since the transform part of a decoder
usually represents a good 30% or more of the work required for the
decode, shaving a lump of that off will translate directly to extended
battery life for the media-consuming device user. Go nuts!
Cheers,
--
Andrew
Reply by Chris●November 16, 20112011-11-16
I'm an audio guy, but one possible application would be texture
recognition in images. This requires a 2-D FFT, and scale that to any
reasonably sized image, and you can see where having a *very* tight
FFT would be critical.
On Nov 14, 9:23�am, "Caspian" <matrose@n_o_s_p_a_m.me.com> wrote:
> >Caspian wrote:
>
> >> Hello DSP pros,
>
> >> I'm currently working on an optimized version of FFT for ARM NEON.
> (FHT,
> >> to be exact.)
>
> >> I'm neither a DSP specialist nor mathematician, and it was very hard
> for
> >> me to understand the way FFT works.
>
> >> Now it's about 50% done, and I have my doubt about this : "What's the
> use
> >> of this? What does it bring when my FFT runs 20times faster than some
> >> unnamed opensource ones on handhelds?"
>
> >If you don't know why performance is important than it's probably not
> important.
> >If you do this work because someone asked you to do it than you must ask
> this
> >person why it pays you to improve performance.
>
> >> I really don't see any point int this. Maybe because I know too little
> >> about DSP.
>
> >Better performance could reduce power consumption.
>
> >bye
> >Andreas
> >--
> >Andreas H nnebeck | email: a...@gmx.de
> >----- privat ---- | www �:http://www.huennebeck-online.de
> >Fax/Anrufbeantworter: 0721/151-284301
> >GPG-Key:http://www.huennebeck-online.de/public_keys/andreas.asc
> >PGP-Key:http://www.huennebeck-online.de/public_keys/pgp_andreas.asc
>
> Usually I'm the one in the company asking for the last bit of performance.
> :)
> I've been working mostly on graphics and image processing where the
> performance increases were literally visible.
>
> But FFT..... I simply know too little about its field of application.
>
> Any idea where heavily FFT'ing large blocks would be necessary on
> handhelds?
>
> cya
>
> PS : Deutsch kann ich auch, sogar besser als Englisch. Die sind aber beide
> Fremdsprachen fuer mich. Vielen Dank.
Reply by Rick Lyons●November 15, 20112011-11-15
On Mon, 14 Nov 2011 17:42:53 +0200, Tauno Voipio
<tauno.voipio@notused.fi.invalid> wrote:
[Snipped by Lyons]
>
>Get a good book, e.g.
>
> Richard G. Lyons, Understanding Digital Signal Processing
>
>read and understand it, and return with follow-up questions
>after that.
Hello Tauno,
if you send an E-mail to me telling me
which Edition, and which Printing Number, of my
book that you have, I'll send you the book's errata.
See Ya',
[-Rick-]
Reply by Andreas Huennebeck●November 15, 20112011-11-15
Caspian wrote:
>[..]
>Andreas wrote:
>>
>>If you don't know why performance is important than it's probably not important.
>>If you do this work because someone asked you to do it than you must ask this
>>person why it pays you to improve performance.
>>[..]
>>Better performance could reduce power consumption.
>
> Usually I'm the one in the company asking for the last bit of performance. :)
> I've been working mostly on graphics and image processing where the
> performance increases were literally visible.
>
> But FFT..... I simply know too little about its field of application.
>
> Any idea where heavily FFT'ing large blocks would be necessary on
> handhelds?
Reply by Andreas Huennebeck●November 15, 20112011-11-15
Rune Allnor wrote:
> On 14 Nov, 15:23, "Caspian" <matrose@n_o_s_p_a_m.me.com> wrote:
>
>> PS : Deutsch kann ich auch, sogar besser als Englisch. Die sind aber
>> beide Fremdsprachen fuer mich. Vielen Dank.
>
> Njaaa.... du b�r vel heller tenke p� at engelsk er
> det mest utbredte spr�ket i det hele tatt. Greit nok
> at *du* kan skrive tysk (og jeg kan lese det, i alle
> fall delvis), men de aller fleste her har engelsk
> enten som sitt morsm�l eller som 1. fremmedspr�k.
>
> Hold deg til engelsk. Da n�r du det st�rste publikum -
> og det st�rste antall potensielle hjelpere.
Reply by Fred Marshall●November 14, 20112011-11-14
On 11/14/2011 6:23 AM, Caspian wrote:
> Any idea where heavily FFT'ing large blocks would be necessary on
> handhelds?
No.
It's entirely up to your application.
You might do FFTs because you need the *output* to display.
You might do them for some other convenience.
You might do them on larger blocks to improve convolutional efficiency.
etc.
In the end, it depends on whether you have enough compute power to get
things done in an *acceptable* amount of time. Well, at least that's
most often the driving consideration. By "compute power" we could mean
time or watts I suppose but usually it's time.
So, you could choose to do a "straight" DFT if the software just comes
to you for free and the performance is adequate. Only you will know.
Or, you could use an FFT if the software just comes to you for free and
the performance is adequate .. for the same size it will be better than
a straight DFT.
etc.
Fred
Reply by Rune Allnor●November 14, 20112011-11-14
On 14 Nov, 15:23, "Caspian" <matrose@n_o_s_p_a_m.me.com> wrote:
> PS : Deutsch kann ich auch, sogar besser als Englisch. Die sind aber beide
> Fremdsprachen fuer mich. Vielen Dank.
Njaaa.... du b�r vel heller tenke p� at engelsk er
det mest utbredte spr�ket i det hele tatt. Greit nok
at *du* kan skrive tysk (og jeg kan lese det, i alle
fall delvis), men de aller fleste her har engelsk
enten som sitt morsm�l eller som 1. fremmedspr�k.
Hold deg til engelsk. Da n�r du det st�rste publikum -
og det st�rste antall potensielle hjelpere.
Rune
Reply by steve●November 14, 20112011-11-14
On Nov 14, 8:22�am, "Caspian" <matrose@n_o_s_p_a_m.me.com> wrote:
> Hello DSP pros,
>
> I'm currently working on an optimized version of FFT for ARM NEON. (FHT, to
> be exact.)
>
> I'm neither a DSP specialist nor mathematician, and it was very hard for me
> to understand the way FFT works.
>
> Now it's about 50% done, and I have my doubt about this : "What's the use
> of this? What does it bring when my FFT runs 20times faster than some
> unnamed opensource ones on handhelds?"
>
> I really don't see any point int this. Maybe because I know too little
> about DSP.
>
> Could someone enlighten me? It would be really great to be motivated
> again.
>
> Thanks in advance!
If you write an FFT that runs 20 times quicker, you can use a
processor 20 times slower = less power, less cost
OR
you can perform 20 times more funtions, = more accuracy, more
features, more stuff to put in your advertising brochure
Reply by Tauno Voipio●November 14, 20112011-11-14
On 14.11.11 4:23 , Caspian wrote:
>> Caspian wrote:
>>
>>> Hello DSP pros,
>>>
>>> I'm currently working on an optimized version of FFT for ARM NEON.
> (FHT,
>>> to be exact.)
>>>
>>> I'm neither a DSP specialist nor mathematician, and it was very hard
> for
>>> me to understand the way FFT works.
>>>
>>> Now it's about 50% done, and I have my doubt about this : "What's the
> use
>>> of this? What does it bring when my FFT runs 20times faster than some
>>> unnamed opensource ones on handhelds?"
>>
>> If you don't know why performance is important than it's probably not
> important.
>> If you do this work because someone asked you to do it than you must ask
> this
>> person why it pays you to improve performance.
>>
>>> I really don't see any point int this. Maybe because I know too little
>>> about DSP.
>>
>> Better performance could reduce power consumption.
>>
>> bye
>> Andreas
>> --
>> Andreas H�nnebeck | email: acmh@gmx.de
>> ----- privat ---- | www : http://www.huennebeck-online.de
>> Fax/Anrufbeantworter: 0721/151-284301
>> GPG-Key: http://www.huennebeck-online.de/public_keys/andreas.asc
>> PGP-Key: http://www.huennebeck-online.de/public_keys/pgp_andreas.asc
>>
>>
>
> Usually I'm the one in the company asking for the last bit of performance.
> :)
> I've been working mostly on graphics and image processing where the
> performance increases were literally visible.
>
> But FFT..... I simply know too little about its field of application.
>
> Any idea where heavily FFT'ing large blocks would be necessary on
> handhelds?
>
> cya
>
>
> PS : Deutsch kann ich auch, sogar besser als Englisch. Die sind aber beide
> Fremdsprachen fuer mich. Vielen Dank.
Get a good book, e.g.
Richard G. Lyons, Understanding Digital Signal Processing
read and understand it, and return with follow-up questions
after that.
--
Tauno Voipio
(Und dasselbe auf Deutsch)