FFT of size 1536

On 04/28/2013 12:23 AM, Vladimir Vassilevsky wrote:
> On 4/27/2013 3:24 PM, glen herrmannsfeldt wrote:
>> Vladimir Vassilevsky <nospam@nowhere.com> wrote:
>>> On 4/25/2013 12:49 PM, glen herrmannsfeldt wrote:
>>>> Vladimir Vassilevsky <nospam@nowhere.com> wrote:
>>>>> On 4/25/2013 11:57 AM, coolboy03 wrote:
>>>>>> I just want to know how can we implement the FFT of size 1536.
>>
>> (snip)
>>
>>>>> For any purpose other then homework, use FFT size of 2048 and pad data
>>>>> with zeroes.
>>
>>>> Unless the data is periodic in 1536 samples.
>>
>>> Then resample 1536 into 4096
>>
>> Hmm. Easier than resampling to 2048 I suppose.
>
> The 2048 would require heavy antialiasing.
>
>> But the radix 3 step isn't that hard to do, is it?
>
> Yes, but I doubt efficiency would be any better.
>
> VLV
>
>
The number of real multiplications of 1536 radix-3 is less than 70% of 
that of the 2048 split-radix.

>On 4/27/2013 3:24 PM, glen herrmannsfeldt wrote:
>> Vladimir Vassilevsky <nospam@nowhere.com> wrote:
>>> On 4/25/2013 12:49 PM, glen herrmannsfeldt wrote:
>>>> Vladimir Vassilevsky <nospam@nowhere.com> wrote:
>>>>> On 4/25/2013 11:57 AM, coolboy03 wrote:
>>>>>> I just want to know how can we implement the FFT of size 1536.
>>
>> (snip)
>>>>>> With reading
>>>>>> the literature I came to know that we do it by splitting 1536 into
three
>>>>>> block of 512 and then take their fft. In the end we combine all the
ffts
>>>>>> and multiply with the respective twiddle factors to and apply the
radix 3
>>>>>> stage. Can anyone help me with the C coding regarding it. Suppose
that I
>>>>>> have done the coding for three 512 sub transforms .
>>>>>> Now I want to go to the radix 3 stage.
>>
>>>>> For any purpose other then homework, use FFT size of 2048 and pad
data
>>>>> with zeroes.
>>
>>>> Unless the data is periodic in 1536 samples.
>>
>>> Then resample 1536 into 4096
>>
>> Hmm. Easier than resampling to 2048 I suppose.
>
>The 2048 would require heavy antialiasing.
>
>> But the radix 3 step isn't that hard to do, is it?
>
>Yes, but I doubt efficiency would be any better.
>
>VLV
>
>
>

Actually FFT of 1536 is used in PRACH LTE 3gpp for a transmission bandwith
of 15Mhz. I have already implemented in matlab, and I can' use any library
or software because I am working in a software company i have to build my
own regardless of the performance of computation.

Am 06.05.13 10:26, schrieb coolboy03:
> Actually FFT of 1536 is used in PRACH LTE 3gpp for a transmission bandwith
> of 15Mhz. I have already implemented in matlab, and I can' use any library
> or software because I am working in a software company i have to build my
> own regardless of the performance of computation.

If you are doing MATLAB, there is no need to twiddle with factors. Just 
call fft(x) which will do it for you (it uses FFTW internally). 
Otherwise kissfft has a very liberal license (that was one reason to 
make it) and can be included in commercial projects without problems.

	Christian