During Summer holidays I experimented about decimation scheme of BreadBoard RF103.
The ExtIO_sddc.dll processes the ADC real signal stream. The sampling rate is 64 MHz.
The output is a decimated I&Q complex signal stream at 16 MHz, 8MHz, 4MHz or 2MHz.
Filtering and tuning are integrated.
The dll uses CyAPI library to connect BBRF103 hardware via USB3.0.
A USBthreadProc thread uses 16 buffers queue to receive data chunk of 65536 samples (short).
One buffer’s time duration is about 1 ms at 64 MHz.
I configured USBthreadProc to run at high priority to be responsive to hardware timing.
The USBthreadProc output buffer is processed by the class RFddc that has an own buffer array of 16 elements.
The time of a circular turn of 16 buffers queue allows the digital down conversion algorithm of each chunk to complete on a separate thread.
The previously computed output vector is returned while the signal processing of the buffer is started. A priority for these threads below normal seems good enough.
Frequency domain signal processing is used. An overlap and add FFT scheme processes the buffer 65536 sample frame and overlap and save scheme glues the buffers together.
The following diagram represents the processing of one buffer frame and it is implemented in every thread of the RFddc pool of 16.
1) It is the input sample array of short. It is a real signal. The frame buffer is a sequence of 65536 samples (it can be seen as a sequence of 64 *1024 slice).
2) The last 1024 slice in the past is copied at the beginning of the array to form a frame of 65 *1024 = 66560 samples. This is the overlap and save scheme.
3) A complex array 66560 samples long is obtained from (2) adding a zero imaginary component.
4) Starting from 0 the array is divided into slices of 768 samples to implement override and add (Overlap and add method) with an overlap of 256 and a fast Fourier transform (FFT) of 1024 sample frame.
5) Every 768 slice is copied into a 1024 one adding a tail of 256 sample zero filled.
85 slice of 1024 complex samples.
6) A FFT forward is applied to the every 1024 slice. 85 slices in frequency domain are computed.
7) For every slide a circular shift of the FFT’s bins is used to implement tuning to the IQ carrier frequency.
The resolution is 64000000/1024 = 62500 Hz. This coarse step is good enough for HDSDR tuning that uses its own fine adjustment. A phase adjustment is required depending on the tuning bin position.
8) A low pass filter of the signal is implemented as fast convolution (https://en.wikipedia.org/wiki/Convolution#Fast_convolution_algorithms) multiplying the FFT bins by the complex conjugate (https://en.wikipedia.org/wiki/Complex_conjugate) frequency response of the filter (Hw*). To use the fast convolution approach the length of the time filter response ht is limited to (1024 -768) +1 = 257 samples.
9) Decimation in frequency is used. The implemented output lengths are 256,128,64,32 that obtains output rate of 16MHz, 8MHz, 4MHz, 2MHz. It is made just copying the decimated FFT bins chunk near zero frequency.
10) The resulting output is a sequence of decimated FFT ( 256,…) bins. Steps (7) (8) (9) are implemented together in a copy and modify loop.
11) The 85 slices of decimated FFT.
12) The FFT inverse is computed.
13) The time output is computed with overlap and add of the 85 slices. The first 256 samples and the latest 512 are dropped using overlap and save (Overlap save method) of the 64 * decimated FFTN samples frames.
The function has an array of 65536 samples input and returns an array of 16.384 samples at 16MHz, 8.192 at 8MHz, 4.096 at 4MHz, 2.048 at 2MHz.
A separate thread processes the signal from each one buffer.
CPU use of HDSDR and ExtIO_sddc, V0.95 ADC 64MHz, IQ 16Msps ; 60 s plot
I made some debug measuring the time jitter of USBthreadProc 16 buffer cycle.
The theoretical time is 16 * 1.024 ms = 16.384 ms.
Here after a plot of the measured duration time - 16.384 ms.
The plot shows that the peak jitter is within +/- 3mS.
USBthreadProc 16 buffer circle timing jitter running at 64 MHz in 16MHz sampling output, 60 s plot.
I named this release version 0.95 and I save it in a separate GitHub repository at:
I switched to the integrated Visual studio Git and it was simpler to me to keep it separate from other BBRF103 project components.
To be continued.