Thread Subject: FFT Quantization

I was wondering if anyone had suggestions on how to vary the number of bits MATLAB uses to compute a windowed FFT. I'm trying to find the "breaking point" for an algorithm and my professor suggested I step the quantization downwards from 16 bits until it no longer works. Any suggestions would be appreciated.

Best regards,
John L.

You might want to create a function that pre-processes your data array
before you pass it to the FFT function. You could use the pre-processor to
"quantize" your data by mapping it to a limited set of specified values, for
example [ 0 ... 64] for 6-bit or [ 0 ... 32 ] for 5-bit, etc.

The declaration might be something like:

   function Y = quantize(X)

where X and Y are arrays of the same size and shape. You might also assume
that X is of class double, and type cast Y so that it is of class uint8.

Does that help?


Rick




"John L" <tturaiderjohn@mathworks.com> wrote in message
news:f77vno$gpg$1@fred.mathworks.com...
>I was wondering if anyone had suggestions on how to vary the number of bits
>MATLAB uses to compute a windowed FFT. I'm trying to find the "breaking
>point" for an algorithm and my professor suggested I step the quantization
>downwards from 16 bits until it no longer works. Any suggestions would be
>appreciated.
>
> Best regards,
> John L.

I guess what would be more appropriate would be an FFT based on a lookup table where I could quantize the lookup table...I'm trying to port this algorithm to an embedded environment...So far I can get down to about an 11 bit representation before things start to fall apart, but I would like to avoid having to make the double precision calculation before I can quantize down. Hopefully that makes sense, I'm somewhat sleep deprived at this point. ;-)