Fast Fourier Transform - Digital Signal Processing and Applications with the C6713 and C6416 DSK - page 240

Several buffers are used, and iobuffer is the primary input/output buffer. At
each sampling interval, the ISR is executed. The next output value is read from
iobuffer , output to the codec, and then replaced by a new input sample. After
PTS/2 sampling instants, iobuffer contains a new frame of PTS/2 input samples.
This situation is signaled by setting flag to 1.
The main program waits for this flag signal using
while (flag == 0);
and subsequently carries out the following operations:
1. Resets flag to 0
2. Copies the contents of the buffer iobuffer (frame of new input samples) to
the first PTS/2 locations of the buffer samples
3. Copies the contents of the buffer overlap (previously computed frame of
output samples) to the buffer iobuffer
4. Processes the new frame of input samples to compute the next frame of output
samples
The frame processing operation (within an infinite loop) has PTS/2 sampling
periods in which to execute and comprises the following steps:
1. The contents of the last PTS/2 locations of the samples buffer (real parts)
are copied to the overlap buffer. These time-domain data may be thought
of as the overlapping latter half (PTS/2 samples) of the previous frame pro-
cessing operation.
2. The last PTS/2 locations of the buffer samples are zero-padded. The buffer
samples now contains PTS/2 new samples followed by PTS/2 zeros.
3. The buffer samples is transformed in-place into the frequency domain using
a PTS-point FFT.
4. The complex frequency-domain sample values are multiplied by the complex
frequency-domain filter coefficients stored in h .
5. The results are transformed back into the time domain by applying a PTS-
point IFFT to the contents of the samples buffer. The resulting PTS time-
domain samples will be real-valued.
6. The contents of the first PTS/2 locations of the buffer samples (i.e., the
former half of the current frame processing result) are added to the contents
of the overlap buffer.
Since the input and output signals are real-valued, so are the buffers iobuffer
and overlap . However, since the frequency-domain representation of these signals
is complex, the buffer samples and the array of filter coefficients h are complex,
requiring two floating-point values (real and imaginary parts) per sample.