ARM7 | BIOS | Bug | EVM | JTAG | Linker | LOG_printf | McBSP | Profiling | Relocation | RTDX | Simulator | Target | Watch

Hello,

Background:
I'm designing a FMCW radar with an embedded C5509A DSP. The radar 
produces a ramped FM signal, which corresponds to beat frequencies 
and thus range. The ramped FM signal is sampled 128 times per period, 
and I take 64 periods. Thus I'm left with a 64x128 (8192 words) of 
sampled data.

Where rfft/cfft come into play:
I need to perform a 2D FFT on the data so I can extract range and 
doppler information. I do the following:

1) Work on rows first. Bit reverse the row's samples. Perform 128-
point rfft across a single row's 128 real samples (columns). This 
produces 64 real and imaginary values per row. I'm using SCALING for 
the rfft.

2) Then work on columns. Perform 64-bit cfft across rows, for each 
column. I'm using SCALING for the cfft. The result is still 64 periods
(rows) of 64 real and 64 imaginary points. 

3) Bit reverse the result.

My questions:
A) Since I'm using scaling in the rfft, do I need to "rescale" my 1st 
FFT results before performing the 2nd FFT? I believe this to be 
necessary because if the scaled results were immediately fed into the 
2nd FFT, the dynamic range would be limited. I would think you'd have 
to multiply the results of the 1st FFT by 2^log_base2(N) where N is 
the number of FFT points. Thus for a 128-point FFT, I'd need to 
multiply the 1st FFT results by 128? (Got this from SPRA948, page 10).

B) Does the input to the rfft/cfft functions need to be in Q15 
format? I think this is the case since the input is type "DATA" which 
is by default Q15 format (I believe).

(You need to be a member of code-comp -- send a blank email to code-comp-subscribe@yahoogroups.com )